July 31, 2018 S. K.

Chaturvedi, REC IIT Kharagpur 1

RELIABILITY
ESTIMATION AND LIFE
TESTING
Introduction
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 2

RE600021:RELIABILITY ESTIMATION AND

LIFETESTING (3-1-0 : 4 Credit)
Parameter Estimation, Regression analysis.
Interval Estimation procedure for exponential,
Gamma, Weibull, Log-normal and Fatigue life
models. Point and interval reliability estimation.
Testing reliability hypotheses for mean of
distribution. Tests for Weibull, distribution,
Reliability testing procedure, types of tests,
accelerated life tests-parametric and
nonparametric methods. Continuously increasing
stress tests.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 3

Reference Books

1. Reliability Eng: Elsayed El. Elsayed, Addison Wesseley

2. Reliability and Maintainability Eng: C. E. Ebling, TMH
3. Reliability in Eng Design: K. C. Kapur and L. R.
Lamberson, John Wily and Sons
4. Applied Life Data Analysis: Wayne B. Nelson: John Wily
and Sons
5. Reliability and Life Testing: Dimitri Kececioglu, Vol I &
II, DEStech Publication.
6. Practical Methods for Reliability Data Analysis: J . I.
Ansell and M . J. Phillips, Oxford Publications
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 4

6. Statistical Models and Methods for Life Time Data, J. F. Lawless,

John Wily and Sons,
7. Methods for Statistical Analysis of Reliability and Life Data: Nancy
R. Mann, Ray E. Schafer, N. D. Singpurwalla, John Wily and Sons
8. The New Wibull Handboo, Robert B. Abernethy
9. Statistical Methods for Reliability Data, William E. Meeker and
Luis Escobar, John Wily and Sons
10. Accelerated Reliability Engineering, Gregg K. Hobbs, Hobbs Eng
Corporation
11. Accelerated Testing-Statistical Models, Test Plans, and Data
Analysis, Wayne Nelson, John Wily and Sons
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 5

PART I
DATA AND DATA TYPES
 6

Reasons for Collecting Data

• Backbone of Reliability Studies.
• Provide valued information to engineers and managers.
• Assessing characteristics of materials.
• Predict product reliability in design stage.
• Assessing the effect of a proposed design change.
• Comparing two or more different manufacturers.
• Assess product reliability in field.
• Checking the veracity of an advertising claim.
• Predict product warranty costs.

• Impossible to have an effective reliability

program without the collection, analysis and
use of information acquired through testing and
operation of products used in industries,
military and consumers’ end.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 7

Why is the assessment and control of product

reliability important?
• Our dependence on continued functioning of
simple to complex systems.

• Failures of some systems can be catastrophic:

injury, loss of life and/or costly lawsuits can
occur.

• More often, repeated failure leads to

annoyance, inconvenience and a lasting
customer dissatisfaction that can play havoc
with the responsible company's marketplace
position.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 8

Why is the assessment and control of product

reliability important?
• Service costs for products within the warranty period
or under a service contract are a major expense and a
significant pricing factor. Proper spare part stocking
and support personnel hiring and training also depend
upon good reliability fallout predictions.

• Shipping unreliable products can destroy a company's

reputation.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 9

Why is the assessment and control of product

reliability important?
• Continual assessment of new product reliability
and ongoing control of the reliability of
everything shipped are critical necessities in
today's competitive business arena.

• Reliability is “quality changing over time”.

• A motion picture instead of a snapshot.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 10

Taxonomy of Data

Failure data is likely to be collected under variety of

differing circumstances varying from carefully
controlled lab experiments to data from field.
Lab Data: More precise, more informative per sample
(physics of failure, mechanism of failure occurrence
etc…)
Field data: represents actual environmental loading
conditions, i.e., temperature, humidity, dirt etc., which
might be difficult simulate in labs.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 11

Taxonomy of Data
Failure data may be classified in several other ways as well:

 Operational versus test-generated data

 Grouped versus ungrouped data
 Large samples versus small samples
 Complete versus censored data
 Other synonyms
• Life data.
• Failure-time data.
• Survival data.
• Event-time data.
• Also, degradation data (somewhat different)
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 12

What are some common difficulties with

reliability data and how are they overcome?
• The Paradox of Reliability Analysis: The
more reliable a product is, the harder it is to
get the failure data needed to "prove" it is
reliable!

• Censoring and Lack of Failures are two

closely related problems that are typical with
reliability data, not common with most other
forms of statistical data. Typically, the
solutions involve making additional
assumptions and using complicated models.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 13

• Censored or Suspended Data

• In many cases when life data are analyzed, all of the units in
the sample may not have failed.
• The event of interest was not observed or
• The exact times-to-failure of all the units are not known.

• Censoring restricts our ability to observe T.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 14

Censoring
• Some sources of censoring are:
• Fixed time to end test (lower bound on T for unfailed units).
• Test terminated after a certain number of failures.
• Inspections times or grouped or readout (upper and lower bounds
on T).
• Staggered entry of units into service leads to multiple
censoring.
• Where units are introduced into the field or test at different times.
• Multiple failure modes (also known as competing risks,
and resulting in multiple right censoring):
• independent (simple).
• non independent (difficult).
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 15

Data Classification
• Complete
• Singly Censored data
• Censored on left

• Censored on right
• Type I censoring
• Type II censoring

• Interval Censored

• Multiply censored data

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 16

Complete Data

• We tested five units

and they all failed and
their times-to-failure
were recorded.
• Complete data is
much easier to work
with than censored
data.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 17

Singly Censored-Left
•A failure time is only
known to be before a
certain time.
• This is identical to interval
censored data in which the
starting time for the interval
is zero.
• we may know that a certain
unit failed sometime before
100 hours but not exactly
when..
• it could have failed any time
between 0 and 100 hours)
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 18

Singly Censored-Right
• Data sets are composed
of units that did not fail.
• The term "right
censored" implies that
the event of interest, i.e.
the time-to-failure, is to
the right of our data
point.
• In other words, if the
units were to keep on
operating, the failure
would occur at some
time after our data point
(or to the right on the
time scale).
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 19

Censoring Type-I
• Right Censored (T Fixed, r random, exact
time to failures are recorded)
• During the T hours of test we observe r failures
(where r can be any number from 0 to n). The
(exact) failure times are t1, t2, ..., tr and there are
(n - r) units that survived the entire T-hour test
without failing.

• This type of censoring is also called "right

censored" data since the times of failure to the
right (i.e., larger than T) are missing.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 20

Censoring-Type II
• Censored Type II Data
• We observe t1, t2, ..., tr, where r is specified in advance. The
test ends at time T = tr, and (n-r) units have survived.
• Again we assume it is possible to observe the exact time of
failure for failed units.
• Type II censoring has the significant advantage that you know in
advance how many failure times your test will yield - this helps
enormously when planning adequate tests.
• However, an open-ended random test time is generally
impractical from a management point of view and this
type of testing is rarely seen.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 21

Censoring-Interval
• Readout /Interval Data /Inspection
• Sometimes exact times of failure are not known;
only an interval of time in which the failure
occurred is recorded.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 22

Lack of failures
• Failure data is needed to accurately assess
and improve reliability - this poses problems
when testing highly reliable parts. it becomes
increasingly difficult to assess the failure rates
of highly reliable components.

• When fitting models and estimating failure

rates from reliability data, the precision of the
estimates (as measured by the width of the
confidence intervals) tends to vary inversely
with the square root of the number of failures
observed - not the number of units on test or
the length of the test.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 23

Lack of failures
• In other words, a test where 5 fail out of a
total of 10 on test gives more information than
a test with 1000 units but only 2 failures.

• Even with an affordable sample sizes and

reasonable time periods, we will have few or
no failures. This gives little or no information
for accomplishing the two primary purposes of
reliability testing, namely:
• accurately assessing population failure rates
• obtaining failure mode information to feedback for
product improvement.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 24

How can tests be designed to overcome an

expected lack of failures?
• Testing at much higher than typical stresses
can yield failures but models are then needed
to relate these back to use stress.
• Testing is ALT and the models that relate high
stress reliability to normal use reliability are
called acceleration models.
 Data Classification: A Note
• Depending on the event that we want to measure, the
data type classification (Comp. or Sus.) is open to
interpretation.
• Under certain circumstances, and depending on the question
one wishes to answer, a specimen that has failed might be
classified as suspended for analysis purposes.
• For Example…

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 25

• The data would be a
complete data, if the objective
is to determine probability of
failure. (Ignoring the mode)
• Failure data due to mode B
and C would be treated as
suspended, if the analysis is
to be done due to failure
mode A only.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 26

 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 27

Distinguishing Features of Reliability Data

• Data are typically censored .
• Positive responses (e.g., time or cycles to
failure) need to be modeled.
• Model parameters not of primary interest
• Instead failure rates, MTBF, quintiles,
probabilities etc…
• Extrapolation often required.
• e.g., want proportion failing at 3 years for a
product in the field only one year.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 28

PART II :
DATA ANALYSIS
• LARGE
• SMALL
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 29

Strategy for Data Analysis, Modeling,

and Inference
1.Model-free graphical data analysis - Non Parametric.
• Begin the analysis by looking at the data without making any
distribution or other model assumption.
• Primary tool-> graphical analysis

2.Model fitting (parametric including possible use of prior information).

• For many applications, it is useful to fit one or more parametric
models for the purpose of description, estimation, or prediction.
• Generally, this process leads to simple to more elaborate models
depending on purpose of study, amount of data and other available
information.

3.Examine appropriate diagnostic for adequacy of model assumption.

• Especially where there is little data.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 30

Strategy for Data Analysis, Modeling,

and Inference
4. Inference: estimation or prediction (for most
reliability application, statistical intervals to
reflect statistical uncertainty/variability).

5. Graphical display of model-fitting results.

6. Graphical and analytical diagnostics and

assessment of assumptions.

7. Sensitivity Analysis

8. Conclusions.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 31

• As stated, analysis begin with tools (graphical/analytical) requiring no

strong model assumptions to avoid distortion that might be caused by
using inadequate assumptions.
• Such estimates may be all that are needed for some cases.
• For others, they are the intermediate steps towards a more precise or
extensive interferences, if the additional assumptions of such models are
valid.
• Non-parametric Approach
• No assumption regarding the distribution.
• Distribution free properties of the data examined.
• Mean, variance etc… are obtained without reference to any
distribution.
 Preliminary Data Analysis
• The purpose of analyzing failure data is to obtain
information on the type of failure distribution, nature
of hazard rate and reliability characteristics of a
component or a unit.
• Since failure density function is a measure of the
rate at which failures occur, it provides useful
information to a designer on how and where the
failures are distributed?
• In the same way, the reliability characteristics tells a
designer how fast the reliability of a unit is decaying
over time or whether or not the designed reliability
goals are met over the specified mission time?
 Preliminary Data Analysis
• The preliminary information on the underlying failure
model can be obtained if we plot the failure density,
hazard rate and reliability functions against time.
• We can define piecewise continuous functions for these
three characteristics by selecting some time intervals t.
• This discretization eventually in the limiting conditions, i.e.,
t0 or when data is large would approach the
continuous function analysis.
• Let us consider that the failure times of a population of 30
electrical bulbs in an underground subway are given in the
table that follows.
Preliminary Data Analysis
 Preliminary Data Analysis
• From the foregoing data, we observe that the
range of failure times is (715.44-3.77) and we
can use Sturges Formula , i.e.,
n = 1 + 3.3 Log10N
where N is the total data points and n is the
optimum number of intervals for developing
histograms.
Therefore for our problem, we have N=30.Thus
we have n=6. However before we group the
failure times in various intervals, it is helpful to
order the failure times such as given in the next
table.
Preliminary Data Analysis
 Preliminary Data Analysis
• Now we are in a position to build the interval table and
calculate R(t), f(t) and h(t). While it is straight forward
to calculate R(t), one can use the following
expressions for calculating f(t) and h(t), which remain
constant during the interval of the histogram.

• Where is the number of failures during the

interval ti and is the number of survivals at
the beginning of the interval ti. Likewise R(t) is also
defined at the beginning of the interval ti.
Preliminary Data Analysis
Preliminary Data Analysis
Preliminary Data Analysis
Preliminary Data Analysis
6 48
7 43
8 38
9 34
10 31
11 28
12 40
13 60
14 75
15 60
16 42
17 15
18 5
19
Example: Plots of Q(t) and R(t)
1 1

0.8 0.8

0.6 0.6

0.4 0.4
Q(t)

R(t)
0.2 0.2

0 0
0 5 10 15 0 5 10 15
Time Interval Time Interval
Example: Plots of f(t) and (t)
0.14 0.5

0.12
0.4
0.1

0.08 0.3

0.06
0.2
f(t)

(t)
0.04

0.02 0.1

0
0 5 10 15 0

Time Interval
0 5 10 15
Time Interval
 Analysis with Censored Data
• To illustrate this case, let us consider that
electronic sensors are provided on 50 equipment
(one to each) to monitor the performance.
• However, if the equipment has failed and the
repair is unsuccessful, the observation on the
sensor attached to that equipment is terminated.
• The observations are continued for 4 years.
Twelve sensors are still found to be functioning.
• The following table provides the computation of
nonparametric reliability estimation of the sensors
including the observations recorded.
 Analysis with Censored Data
ns  1  nf m
R (ti )  
i i

i 1 ns i  1
i ti ns i  1  nf i R(ti)
ns i nf i nti ns i  1 nsi 1  nfi
months ns i  1

 nf i  29  nt i 9
Analysis with Censored Data
If we assume exponential failure distribution for the time to
failure of the sensors, the MTTF is given by:

=(522+201+576)/29 = 44.793 months.

= 3.732 years.
The theoretical and observed reliability functions are shown
in the following slide.
Non-parametric Estimation of Reliability
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 50

Empirical Distributions
• Calculated directly from sample
• Advantage: Avoids uncertainty associated with selecting
and fitting distributions
• Disadvantage: extrapolation and sample may not be
representative
• [min, max] - given by data, estimated by extending the
tails, or inferred by judgment
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 51
Large Small
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 52

Non-Parametric Ungrouped Complete Data

• To make inference, based on the information

contained in a sample .

• For small sample, method of ranking

statistics is often employed, where CDF, in
place of pdf is estimated.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 53

Plotting positions (CDF)

ia
• General F(ti)  100 
n  1  2a
where a that varies [0, 0.5]
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 54

The Ranking of the data may be interpreted by

one of the following three assumptions:
(1) The Equal Rank Assumption
 t i ) = 1- R(
F(  ti ) = i
n

(2) The Mean-Rank Assumption

 i
F( t i ) =
n+1
(3) The Median-Rank Assumption
 ti ) = i-.3
F(
n+.4
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 55

Example various ranking schemes for the

assessment of CDF
• Proof-tests of 14 turbo engines provide the time-
to-failure data (in hours): 103; 113; 72; 207; 82;
97; 126; 117; 139; 127; 154; 127; 199; 159.

• The sample size is only N=14, histogram

construction?

• The 14 data points are first Ranked in ascending

order as follows:
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 56

Ranking i Time to fail ti F(i)=i/n F(i)= i/n+1 F(i)=(i-

0.3)/(n+0.4)
1 72 0.07 0.0667 0.0486
2 82 0.14 0.13 0.12
3 97 0.21 0.20 0.19
4 103 0.29 0.27 0.26
5 113 0.36 0.33 0.33
6 117 0.43 0.4 0.39
7 126 0.5 0.47 0.47
8 127 0.57 0.53 0.54
9 127 0.64 0.6 0.6
10 139 0.71 0.67 0.68
11 154 0.78 0.73 0.74
12 159 0.86 0.8 0.81
13 199 0.94 0.87 0.89
14 207 1.0 0.93 0.95
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 57

Empirical PDF & Hazard Rate

 t i+1 ) - R(
R(  ti ) 1
 = -
f(t) = for t i < t < t i+1
t i+1 - t i ( t i+1 - t i )(n+ 1)


f(t) 1
 (t) = = for t i < t < t i+1

R(t) ( t i+1 - t i )(n+ 1- i)

 n+ 1 - i
R( t i ) =
n+ 1
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 58

Sample Mean & Variance

n

MTTF = i=1
ti
n
n 
( t i - MTTF )2
2
s = 
i=1 n -1

2

i=1
2
t
i - n 
MTTF
2

s =
n -1
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 59

Example
Calculation of descriptive Statistics
and Empirical curves

Given the 10 failure times in hours.

Estimate R(t), F(t), f(t), h(t) and MTTF.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 60
Ranked F(t) R(t) = f(t)= h(t) Refer Figure
data (Time) (Median (1-F(t)) 1/(n+0.4)*(ti =f(t)/R(t) 2, for
Rank) +1-ti) Formulae
0 0 1 0.0061 0.0068
15.697 0.0962 0.9038 0.0236 0.0292
19.776 0.1923 0.8077 0.5721 0.8040 MTTF =
40.7995
19.944 0.2885 0.7115 0.0220 0.0358
24.307 0.3846 0.6154 0.0169 0.0326
29.992 0.4808 0.5192 0.0237 0.056 STD dev. =
24.5545
34.051 0.5769 0.4231 0.0062 0.0189
49.627 0.6731 0.3269 0.0280 0.1212
53.066 0.7692 0.2308 0.0043 0.0323 Median =
32.0222

75.198 0.8654 0.1346 0.0086 0.2245

86.332 0.9615 0.0385
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 61

Curves for R(t), f(t) and h(t)

Empirical Reliability Curve Failure Distribution (pdf) Hazard Plot, h(t)
1 0.7 0.9

0.9 0.8
0.6
0.8
0.7
0.7 0.5
0.6
0.6
0.4 0.5
R(t)

h(t)
0.5
f(t)

0.3 0.4
0.4
0.3
0.3 0.2
0.2
0.2
0.1
0.1 0.1

0 0 0
0 50 100 012345678910
1112131416
511718290 012345678910
1112131416
511718290
Time Interval Interval
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 62

Grouped Complete Data

Let n1, n2, ..., nk be the number of units having

survived at ordered times t1, t2, ..., tk respectively.

 ni
R( t i ) = , i = 1, 2,  , k
n
 t i+1 ) - R(
R(  ti )
 = -
f(t) for t i < t < t i+1
( t i+1 - t i )

ni - ni+1
=
( t i+1 - t i )n
Animated
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 63

Grouped Complete Data


f(t) ni - ni+1
 (t) = = for t i < t < t i+1

R(t) ( t i+1 - t i ) ni

k -1
( ni - ni+1 ) ( t i + t i+1 )
MTTF =  t i
 where t i =
i=0 n 2

( n -n ) 
k -1

t
i i+1
2
s =
2
i - MTTF 2

i=0
n
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 64

Example
Seventy compressors are observed at 5 month intervals with the following
number of failures: 3, 7, 8, 9, 13, 18, and 12.

UPPER NBR NUMBER RELI- FAILURE HAZARD

BND FAIL SURVIVE ABILITY DENSITY RATE

0 0 70 1.000 .0086 .0086

5 3 67 .957 .0200 .0209
10 7 60 .857 .0229 .0267
15 8 52 .743 .0257
67 .0346
20 9 43

R(5)
.614 = .0371 = 0.957.0605
25 13 30 .429 70
.0514 .1200
30 18 12 .171 .0343 .2000
35 12 0 0.000
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 65

Example

UPPER NBR NUMBER RELI- FAILURE HAZARD

BND FAIL SURVIVE
67 - 60 ABILITY DENSITY RATE
 =
f(t) = 0.0200 for 5 < t < 10
0 0 70 (10 - 5)70 1.000 .0086 .0086
5 3 67 .957 .0200 .0209
10 7 60 .857 .0229 .0267
15 8 52 .743 .0257 .0346
20 9 43 .614 .0371 .0605
25 13 30 .429 .0514 .1200
30 18 12 .171 .0343 .2000
35 12 0 0.000
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 66

Example

UPPER NBR NUMBER RELI- FAILURE HAZARD

BND FAIL SURVIVE ABILITY DENSITY RATE

0 0 70 67 - 60
 (t) = 1.000 .0086= 0.0209 .0086
for 5 < t < 10
5 3 67 .957 .0200
(10 - 5)67 .0209
10 7 60 .857 .0229 .0267
15 8 52 .743 .0257 .0346
20 9 43 .614 .0371 .0605
25 13 30 .429 .0514 .1200
30 18 12 .171 .0343 .2000
35 12 0 0.000
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 67

Example

 [2.5 ( 3) +7.5 (7 )++32.5 (12 ) ]

MTTF = = 21.357
70
2 [2. 5
2
( 3 ) +7. 5
2
(7 )+ +32. 5
2
(12) ]
s = - 21. 357 2 = 76.551
70
or s = 8.75
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 68

Example
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 69

Example
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 70

Non- Parametric Ungrouped Censored Data

• The Formulae for Descriptive Statistics would not be
valid.
• Better to fit a Theoretical Distribution in Conjunction with
Analytical Approaches such as MLE or LSE.
However,
• Graphical analysis of Singly Right Censored Data, may be
computed using the approach similar to the uncensored
data.
• Curves are truncated to the right at the time of the test
termination.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 71

Singly Censored Data

Twenty units are placed on test for 72 hours with the
following failures times recorded:
1.5, 3.2,11.7, 26.4, 39.1, 56.0, 61.3
^
TIME
0
RELIABILITY
1
CDF
0
R(ti )  1  i / (20  1)
1.5 .952381 4.761904E-02
3.2 .9047619 9.523809E-02
11.7 .8571429 .1428571
26.4 .8095238 .1904762
39.1 .7619048 .2380952
56 .7142857 .2857143
61.3 .6666667 .3333333

Cannot compute a mean or variance

Animated
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 72

Multiply Censored Data

A sample consists of a set of ordered failure times
plus censored times:

t1, t2, t3+, … ti, ti+1+, ..., tn

• Product Limit Estimator
• Kaplan-Meier
• Rank adjustment
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 73

Rank adjustment method

• The best practice employed for the analysis of
multiply censored data .

• Censored unit has a probability of failing before or

after the next failure or failures, thereby
influencing the rank of subsequent failure .
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 74

• The adjusted rank is determined by:

Reverse Rank  Previous Adjusted Rank  N  1

Adjusted Rank , i 
Reverse Rank  1
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 75

Example

Analysis for Multiply Censored Data for

Empirical Curves

• The data represents the failure and censored

times in operating day of a turbine vanes.
Calculations for adjusted ranks and F(t) using
median ranking statistics are shown.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 76

Rank Time Reverse Adjusted Rank F(t) = i-0.3/8.4

Rank
1 10+ 8 - -
2 30 7 {7*0 + (8+1)}/(7+1) 0.0982
=1.125
3 45+ 6 - -

4 49 5 {5*1.125+9}/ 6 = 0.255
2.438
5 82 4 {4*2.438+9}/ 5 = 0.411
3.750
6 90 3 5.063 0.567
7 96 2 6.375 0.723
8 100+ 1 - -
Suspended items do not affect rank numbers until
after they occur.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 77

Fitting Theoretical Distributions

• Empirical models do not provide information beyond the range
of the sample data.
• A sample is only a small (random) subset of the population of
failure times, and it is the distribution the sample came from and
not the sample itself which we want to establish.
• Often the failure process is a result of some physical
phenomena which can be associated with a particular
distribution.
• Small sample sizes provide very little information concerning the
failure process. However, if the sample is consistent with a
theoretical distribution, then much "stronger" results are
possible based upon the properties of the theoretical
distribution.
• Use can be made of the theoretical reliability model in
performing more complex analysis of the failure process.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 78

PARAMETRIC APPROACH
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 79

• Preferred Approach
• Encompasses both, i.e., Choices of Distributions and their
parameter estimation.
• Extrapolation beyond the censored data points is
possible.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 80

Parametric Probability Distributions

• Several distribution functions for the pdf (or CDF) have
been employed in various applications

• Their forms are explicit mathematically

• Each may contain one or more free parameter(s)

• The properties of these functions (e.g. the mean,

variance, skew ness, etc.) expressible in terms of their
free parameter(s).
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 81

Candidate Distributions

Category CDF/Characteristics Comments/Applications

Exponential F(t) =1-exp(-t), R(t) Widely used failure

= exp((-t), h(t) =  distribution. Appropriate
f(t) =  exp((-t),  = whenever failures occur
1/, median = randomly and are not age
0.6931/, 2= 1/2, dependent. Used to model
coeff. of Skewness = failures of electronic and
2, peakedness = 6, electrical parts.
1/, called the scale
parameter.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 82

Exponential PDF
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 83

Category CDF/Characteristics Comments/Applicati

ons
Weibull F(t) = 1-exp(-t/) , R(t) = Extensively used due
exp(-t/) , h(t) = to its flexibility, used to
(/)(t/) -1, f(t) = model failure times of
(/)(t/)-1 exp(-t/) ,  electron tubes,
= (1+1/), median =
(0.6931/), 2=
capacitors, ball
2[(1+2/)-{(1+1/}2], bearings, leakage from
Skewness = 3/3 , battery and so forth.
peakedness = 4/4-3, Extended to weibayes
where m=mj=0…m (-1)j analysis for small
mC ((m-j+)/) j
j sample size and even
((1+)/). to analyze situations
, shape parameter, , with no failure.
scale parameter or
characteristic life. 0<<1,
 = 1 (exponential), =2,
(Rayleigh), >3 (approx.
Normal), i.e., it can model
a variety of distributions by
varying .
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 84

Weibull PDF & FR

Shape is determined by .
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 85

Category CDF/Characteristics Comments/Applications

Normal 1
t
1 t    2 No closed form for F(t),
F (t )  
 2  
exp( 
2

 
 ) dt
 R(t), h(t). Standard Normal
tables are employed for
calculations, which has
It can be associated mean=Skewness=
with a failure rate peakedness = 0, and
function that is of variance = 1
wear-out mode. h(t)
rises sharply as t
exceed the MTTF.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 86

Normal PDF
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 87
Category CDF/Characteristics Comments/Applications

No closed form. Derived from

normal distribution. A
Lognormal 1
t
1  ln t  2 competitor of Webull
F (t ) 
st 2
 exp( 
2s 2  t med
) dt
  distribution. This has been used
to model the failure times of
Mean =tmed* exp(s2/2) and
Variance
many semiconductor devises,
=tmed2exp(s2){exp(s2)-1} viz., diodes, LEDs, lasers etc… .
Behaviour as that of left
skewed function. Skewness
dictated by s. s1, the pdf
becomes an exponential,
s0.1, pdf approaches
normal. It can be associated
with constant or wear out
mode. This distribution is
only defined for positive
values of t and more
appropriate than normal.
Like Weibull, it can also take
variety of shape
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 88

Lognormal PDF

Shape is primarily determined by . Spread is primarily

determined by the parameter .
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 89

Graphical Assessment of Distribution and

its Parameters
Objective: To confirm whether it is reasonable to
assume that a given set of data comes from some
hypothesized distribution, e.g. normal, exponential
etc.
One method: Probability plotting.
The idea: Construct special graph “paper” by scaling
the axes so that if the data actually come from the
hypothesized distribution, a reasonably straight line
will fit the data when plotted on the graph.
Let’s see what this means.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 90

Probability Plots
• Probability plotting is a graphical method for
determining whether sample data conform to a
hypothesized distribution based on a subjective
visual examination of the data.

• Probability plotting typically uses special graph

paper, known as probability paper, that has been
designed for the hypothesized distribution.
Probability paper is widely available for the
normal, lognormal, Weibull, and various chi-
square and gamma distributions.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 91

Typically, a cdf looks like the graph below. It’s certainly not a
straight line.

1.0
0.9
0.8
0.7
F (x) 0.6
0.5
0.4
0.3
0.2
0.1
0.0

0 1 2 3 4 5 6 7 8 9 10

x
Now “distort” (transform) the axes so that the
cdf does plot as a straight line.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 92

We won’t go into detail on how to linearize cdfs, but

let’s see how it’s done with the exponential.

Example : Consider the exponential with cdf

FX  x   1  e  x , x  0.
Clearly,
 ln 1  FX  x     x, x  0.

Now construct a – ln(1 – y) scale for the y-axis instead

of the regular linear axis and retain the regular linear
axis for the x-axis. Now you’ve got exponential “paper.”
For a given , plot FX on the transformed y-axis vs. x
and you’ll get a straight line with slope .
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 93

All probability “papers” employ a similar

procedure. Not long ago, it really was paper.
Now we can use software.

O.K. – so now we’ve got a graph with the axes

transformed by some functions g and h so that
g  FX  x    ah  x   b,

Note: For the exponential,

g  y    ln(1  y ), h  x   x, a   , and b  0.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 94

Summary of the Transformation for Plotting

Distribution y = ax + b X-Y plot data (if not Inference

on Probability
Paper)

Exponential 1 y-axis (ordinate) Slope of the

log( )  t
1  F (t ) Logarithmic scale line gives the
parameter ‘λ’
and ‘b’ = 0
x-axis (abscissa)
Linear scale
Weibull 1
y-axis (ordinate Slope of the
log(log(
1  F (t )
))   log(t )   log( )
Logarithmic scale line gives the
parameter ‘’
x-axis (abscissa)
Logarithmic scale
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 95

Distribution y = ax + b X-Y plot data (if not Inference

on Probability
Paper)

Normal Both linear Slope of the

line gives the
 1

1
[ F (t )]  t  parameter
  ‘1/σ’ and ‘b’
= -μ/σ

Lognormal
y 
1
[ F ( t )]
1 

1
[ F (t )]  log(t ) 
 
x  log( t )
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 96

Procedure of Plotting
Objective: to fit a line (linear regression) of the type y =
ax +b to a set of transformed data.
1. Calculate F(ti) : use ranking statistics and plot (t, F(ti))
2. Visual Check: If the points plotted fall in a straight
line, it indicates that the empirical data could be
represented by particular distribution.
3. Fitting the Best Line :The chosen line should be the
best representative of all point.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 97

Therefore,
• Plotting { ti, F(ti)} on the probability paper
would graph as an approximate straight line
• Provides the visual test of a distribution in
comparison to a histogram.
• Gives an initial assessment of distribution’s
parameters involved.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 98

Advantages of probability plotting

Simple, graphical, works better than histograms, especially for
small samples. Less subjectivity than histograms too because no
grouping of the data is required.
Some rules of thumb:

 Some ill fitting in the tails is common, so don’t be too

concerned if this happens.
 Is the straight line fit reasonable? Sometimes you’ll be lucky,
but don’t expect great fits with very small samples.
 There will be some subjectivity, e.g., how good a fit is
“reasonable”? Use an imaginary “fat pencil” test.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 99

Note
 A least square fit to the data is recommended over a
manual plot of the data on probability paper.

 It is more accurate and less subjective than fitting a
straight line by eye.

 An index of fit can also be computed to measure the

fitness of the distribution.

 Many Statistical Packages are available with such

capability
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 100

Example

Use of the various probability plots to

ascertain the distribution for the data as
given in Example 2 .
15.697 19.776 19.944 24.307 29.992 34.051
49.627 53.066 75.198 86.332

• The demo version of WinSmithWeibull™ is

used to obtain the various plots.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 101

Normal Probability Plot

From the point of ordinate, F(t) = 0.5, draw a

horizontal line, which intersects at some point
on the best fit line of the empirical data. Draw a
line vertically downward to the x-axis, where
the numerical value of  can be read. The
numerical value of, , would be equal to the
difference between the values of abscissa, whose
ordinates have values 0.5 and 0.16, or 0.84 and
0.5, i.e.,  = t0.5-t0.16 = t0.84-t0.5. (Note: F(t=) =
0.5)
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 102

Exponential Probability Plot

From the point of ordinate, F(t) =

0.632, draw a horizontal line, which
intersects at some point on the best fit
line of the empirical data. Draw a line
vertically downward to the x-axis,
where the numerical value of  can be
read. (Note: F(t=)=1-exp(-/) =
0.632)
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 103

2-Parameter Weibull Plot

Draw a vertical line from the point of

intersection of the best fit line with
ordinate, F(t) = 0.632, and read the
value for , on the horizontal axis.
(Note: F(t=) = 1- exp(-1) = 0.632)
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 104

Lognormal Probability Plot

Follow the same procedure as that for the

normal distribution. The point of interest
are the numerical values of the horizontal
axis, which correspond to numerical values
of the intersection of the best fit line and the
cumulative probabilities of 0.16, 0.5 and
0.84. Once these values are determined, the
distribution parameter can be obtained as 
= ln (t0.5),  = ln {1/2(t0.5/t0.16 + t0.84/t0.5)}.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 105

3-Parameter Weibull Plot

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 106

2-Parameter Exponential Plot

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 107

Use this when,

• Plot shows concave curvature.
• There should be explanation for why failure can not occur
before certain time.
• A prior knowledge from earlier experience that a third
parameter is appropriate or Larger sample size is
available
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 108

Pitfalls and Cautions

Probability plotting is a very useful descriptive
and analytical method, which is used
extensively in industry. But –
• One distribution may fit a set of data about as well as another. If
inferences are made within the range of the data, the choice won’t
matter much, but extrapolation beyond the range of the data can
lead to large differences.

• Sometimes an engineer’s experience will provide guidance

concerning which distribution to use. For example, some kinds of
random phenomena tend to be lognormal, some Weibull.
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 109

Example : A component used in a power supply was life

tested at 80oC and 120oC. Since the normal operating
temperature is 25oC, this is called an accelerated test.
80oC 120oC
4983 1151
2259 3181 Units are in hours. The
6086 8478 probability plots and
10313 619 parameter estimates appear
2542 684
on the next slide.
6460 3930
8515 2168
Does the data appear to be
9656 235 from a Weibull distribution?
4622 562
9178 1303
2039 1423
15891 1720
1263 4805
805 3263
504 2988
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 110
Weibull Probability Plot for 120 DEG C...80
DEG C
99 120DEGC
80DEGC
95
90
80
70
Parameter
Estimates
60
50
40
Percent

30
20
1200C:
10 shape 1.22780
5 scale 2612.56
3
2

1
800 C:
shape 1.2946
100 1000 10000
Time to failure scale 6131.95

The slope of the lines is . Thus, the steeper the slope,

the higher  is. Here the lines are parallel. The 800C line
is to the right of the 120oC line, indicating a larger  .
Conclusions? Does temperature affect life? How?
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 111

Example : Consider the accelerated test data of previous

Example. Lognormal plots appear below. Can you see much
difference in the fit? Which fit (Weibull or lognormal) fits
better?
Weibull ProbabilityPlot for 120DEGC...80
Lognormal ProbabilityPlot for 80DEGC...120DEGC DEGC
120DEGC 99 120DEGC
99
80DEGC 95 80DEGC
90
95
80
90 70
60
80 50
40
70

Percent
30
Percent

60
50 20
40
30 10
20
5
10
3
5
2

1 1

100 1000 10000 100 1000 10000

Timetofailure Timetofailure
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 112

Example: A fatigue-fracture test on cylindrical specimens of a

particular carbon steel tested at  37.1kg/mm2 stress
amplitude yields the data shown below. Units are in
thousands of load oscillations.

40 38 113 85 304 67
282 157 34 83 42 91
1711 155 273 18 67 24
759 492 738 24 146 67
29 60 52 159 213 179
246 82 112 20 34 34
216 118 117 514 123 87
675 95 150 190 51 89
106 7 92 62 201 46
121 105 74 26 54 26
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 113

Below is a histogram of the data. Given a sufficiently large

random sample, the histogram ought to look reasonably like the
pdf of the distribution whence the sample comes.

Would you say the sample comes from a normal distribution?

40

30
Frequency

20

NAH!!!
10

0 250 500 750 1000 1250 1500 1750

Thousands of Load Oscillations
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 114

Normal Probability Plot for C1

99 ML Estimates

Mean: 172.917
95
StDev: 260.052
90

80
70
Percent

60
50
40
30
20

10
5

0 500 1000 1500

Data
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 115

Now transform the original data by taking natural logs. Below

is a histogram of the log-transformed data.

Now would you say that the (transformed) data come from a
normal distribution?

20
YEAH!!!
So – we can
reasonably say that
Frequency

10
the (original) data
come from a log-
normal distribution.
0

2.0 2.8 3.6 4.4 5.2 6.0 6.8 7.6

(ln) Thousands of Oscillations
 July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 116

Here’s the lognormal probability plot. Conclusions?

Lognormal Probability Plot for Fatigue-Fracture Test

99 ML Estimates

Location: 4.56693
95
Scale: 1.02871
90

80
70
Percent

60
50
40
30
20

10
5

10 100 1000
Cycles-to-fracture
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 117

SOME WPP & THEIR

INTERPRETATONS
 118


July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 119

?
• Indication of a Mixture of FMs.
• Chk Broken Parts and failure reports.
• Try to categorize in groups,
• Redo analysis for each group to
• get a better fit. Otherwise try for
• Other distribution.
•
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 120
?

• Indication of interval or coarse data.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 123

• Curvature indicates lognormal/

• or third Parameter.
• Curvature indicates lognormal/
• or third Parameter.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 125

• Steep beta. Laboratory tests data often have

• Due to their controlled env. And precision in
• recording data.
• 3-P Weibull/lognormal or a distributional analysis.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 126

• 3-P Weibull/lognormal or a distributional analysis.

July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 127

Conclusions
• Identifying candidate distribution is both an art and
science.
It Requires:
• Statistical analysis of data
• An understanding of failure process knowledge of
characteristics of theoretical distribution
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 128

One Suggested Approach

1. Construct a histogram of the failure times,

if possible. The literature advises for
Weibull analysis for the sample size<20.
2. Compute descriptive statistics.
3. Analyze the empirical failure curves.
4. Use prior knowledge of the failure process.
5. Use properties of theoretical distributions.
6. Construct a probability plot.
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 129

PART III

Preliminary Data Analysis

of
Repairable Systems
7/31/2018

Graphical methods
• Ask is there any trend in their times to failure ?
• A simple example : suppose 3 system are observed until time till 12th failure :
7/31/2018 S. K. Chaturvedi, REC IIT Kharagpur 131

Duane plots
• A scatter of Log (N(ti )) vs Log (ti )on a double log paper.
• This N(ti ) vs ti is often called as Cumulative failure rate .
• Duane plots should be roughly linear if power law process ( time between
failures governed by Weibull ) governs the failure times .
• Duane plot of a radar : What do u conclude ?
• The downward slope of the points for sys would suggest reliability improvement
• An upward slope with increase in points would suggest deterioration and if
points roughly lie on a straight line horizontal , its shows stability
July 31, 2018 S. K. Chaturvedi, REC IIT Kharagpur 132
 S. K. Chaturvedi, REC IIT Kharagpur
7/31/2018 133

DUANE PLOT OF RADAR DATA

duane plot
0
0 5 10
-1

-2
ln(N(ti)/ti)

-3 Series1
Linear (Series1)
-4

-5

-6 y = -0.5327x - 1.1934
ln (ti) R² = 0.9097
 S. K. Chaturvedi, REC IIT Kharagpur
7/31/2018 134

TTT PLOTS
• TOTAL TIME ON TEST PLOTS
• PLOT OF Uj vs j/n
• PLOT OF A RADAR : REVEALING IT FOLLOWS A POWER LAW PROCESS

1.2 TTT PLOTS

1
0.8
0.6
Series1
uj

0.4 Linear (Series1)

0.2
0
-0.2 0 0.5 1 1.5
j/n