Fitting parametric distributions

Dr Jane Marshall
PEUSS
 Fitting parametric distributions
Understanding censoring structures
Repaired and non repaired items
Probability plotting
– Weibull plotting
– Estimating parameters
Hazard plotting
– Weibull hazard plotting
– Estimating parameters

2
 Censoring structures
Complete data
Single censored
– Units started together and data analysed before all units have
failed
– Right, interval and left
Time censored
– Censoring time is fixed
Failure censored
– Number of failures is fixed
Multiply censored
– Different running times intermixed with failure times – field data

3
 Complete Data

4
 Single right censored (time)

Item 5 O
Item 4 X
Item 3 O
Item 2 X
Item 1 X

time X Failure event

O Censored event

5
 Single right censored (failure)

Item 5 O
Item 4 X
Item 3 O
Item 2 X
Item 1 X

time X Failure event

O Censored event

6
 Interval Censored

7
 Multiply right censored (time)

Item 5 O
Item 4 X
Item 3 O
Item 2 X
Item 1 X

time X Failure event

O Censored event

8
 Left censored

X Failure event
O Censored event

9
 Repaired and non-repaired data
Non-repaired data when only one failure
can occur and interested in time to failure

Repaired data when interested in the

pattern of times between failures

10
Probability plotting
 Areas to be covered
Introduction to probability plotting
Assumptions
How to do a Weibull plot
Estimating the parameters
Testing assumptions
Examples

12
 What is probability plotting?
Graphical estimation method
Based on cumulative distribution function CDF or
F(t)
Probability papers for parametric distributions,
e.g. Weibull
Axis is transformed so that the true CDF plots as
a straight line
If plotted data fits a straight line then the data
fits the appropriate distribution
Parameter estimation

13
 Assumptions
Data must be independently identically
distributed (iid)
– No causal relationship between data items
– No trend in the time between failures
– All having the same distribution
Non-repaired items
Repaired items with no trend in the time
between failures
Time to first failure of repaired items

14
 Example of test for trend
Machine H fails at the following running
times (hours):

– 15, 42, 74, 117, 168, 233, and 410

Machine S fails at the following running

times (hours):

– 177, 242, 293, 336, 368, 395, and 410

15
 Trend Analysis
machine H running times to failure
machine S running times to failure
8 8
7 7

order num ber

order number

6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 100 200 300 400 500 0 100 200 300 400 500

machine time to failure machine time to failure

This system is getting better with This system is getting worse with
time, the failure times are getting time, the failure times are getting
further and further apart closer and closer together.

In neither case can Weibull analysis be used as there

is trend in the data.

16
Making a Weibull plot
 Rank the data
Probability graph papers are based on plots of
the variable against cumulative probability
For n< 50 the cumulative percentage
probability is estimated using median ranks
tables
For n< 100 use benard’s approximation for the
median rank ri
ri = i - 0.3
n+0.4
Where i is the ith order value and n is the sample size

18
 Example
Failure number Ranked hrs at failure Median Rank from tables
(i) (ti) Cumulative % Failed at ti -
F(t)
1 300 6.7
2 410 16.2
3 500 25.9
4 600 35.5
5 660 45.2
6 750 54.8
7 825 64.5
8 900 74.1
9 1050 83.8
10 1200 93.3

19
 How to draw a
Weibull plot
• Plot times on x-axis
• Plot CDF on y-axis
• Fit line through the data
• Draw perpendicular line from
estimation point to the
fitted line.
• Read off the estimate of β
• η is the value given on
from the intersection line

20
 Interpreting the plot
If the data produced a straight line then:
– The data can be modelled by the Weibull distribution.
If β<1 then data shows a decreasing hazard function
– e.g. Infant mortality, weak components, low quality
If β=1 then data shows a constant hazard function
– e.g. useful life of product
If β>1 then data shows a increasing hazard function
– e.g. wear-out, product reaching end of life
η is the value in time by which 63.2% of all failures
will have occurred and is termed the characteristic life

21
 Bath-tub curve and the Weibull
Hazard function

Useful Life

=1
<1
>1

Time
Infant Wear Out
Mortality
22
 Interpreting the plot
If the data did not produce a straight light
then:
– There may be an amount of failure-free time
• This may appear concave when viewed from the bottom
right hand corner of the sheet
– There may be more than one failure mode present
• This may appear convex shape or cranked shape (also
known as dog-leg shape)
• In this case the data needs to separated into failures
associated with each failure mode using expert
judgement and analysed separately

23
 Example: Poor fit due to 3 months offset

Weibull Plot of Time-in-Service Weibull Plot of Time in service (Months)

2-Parameter Weibull - 95% CI 3-Parameter Weibull - 95% CI
Censoring Column in Censoring - ML Estimates Censoring Column in Censoring - ML Estimates
99 99
90 Table of Statistics 90 Table of Statistics
80 80
70 Shape 1.87010 70 Shape 0.873095
60 60
50 Scale 57.6561 50 Scale 297.337
40 40
30 30
20 Mean 51.1897 20
Thres 2.9997
10 10
5 5
Percent

Percent
3 3
2 2
1 1

0.01 0.01
0.1 1.0 10.0 100.0 0.0001 0.0010 0.0100 0.1000 1.0000 10.0000 100.00001000.000010000.0000
Time in service (Months) Time in service (Months) - Threshold

The same data plotted with a three-Parameter Weibull distribution

shows a good fit with 3 months offset (location – 2.99 months)

24
 Example of two failure modes

Mode 2
Beta =
11.9
Weibull CDF

Mode 1
Beta =
0.75

Time to failure ( hours)

25
 Adjusted rank for censored data

Reverse Median
Rank Time Status rank Adjusted rank rank
1 10 Suspension 8 Suspended...
2 30 Failure 7 [7 X 0 +(8+1)]/ (7+1) = 1,125 9,8 %
3 45 Suspension 6 Suspended…
4 49 Failure 5 [5 X 1,125 +(8+1)]/ (5+1) = 2,438 25,5 %
5 82 Failure 4 [4 X 2,438 +(8+1)]/ (4+1) = 3,750 41,1 %
6 90 Failure 3 [3 X 3,750 +(8+1)]/ (3+1) = 5,063 56,7 %
7 96 Failure 2 [2 X 5,063 +(8+1)]/ (2+1) = 6,375 72,3 %
8 100 Suspension 1 Suspended...

26
 Weibull Analysis using software
tools
Number of software packages that can do
Weibull plotting (and other distributions), these
include:
– Minitab
– Relex
– WinSMITH
– Reliasoft
Concentrate on getting good quality data,
correct assumptions and correct interpretation
from the software

27
Hazard Plotting
 Contents
Assumptions
Fitting parametric distributions
Estimating parameters
Using results for decision making

29
 Assumptions
Non-repaired items
Repaired items with no trend in the time
between failures
Time to first failure of repaired items
Individual failure modes from items
Can easily deal with censored data
in particular multiply censored data

30
 Hazard Plotting
Cumulative hazard function

Relationship allows derivation of

cumulative hazard plotting paper

31
 Weibull Hazard Plotting
and
If H is the cumulative hazard value then

Weibull hazard paper is log-log paper

The slope is 1/  and when H=1, t= 

32
Hazard plotting procedure
Tabulate times in order
For each failure, calculate the hazard interval
– Hi = 1/ no of items remaining after previous
failure/censoring (as estimated by 1/reverse rank)
For each failure, calculate the cumulative
hazard function
– H = H1 + H2 + ………. + Hn
Plot the cumulative hazard against life value

33
 Example 1 – vehicle shock absorbers
Distance (km)
6700 F 17520 F
6950 17540
7820 17890
8790 18450
Distance to failure for 9120 F 18960
9660 18980
Shock absorbers 9820 19410
F denotes failure 11310 20100 F
11690 20100
11850 20150
11880 20320
12140 20900 F
12200 F 22700 F
12870 23490
13150 F 26510 F
13330 27410
13470 27490 F
14040 27890
14300 F 28100

34
 Example 1 – vehicle shock absorbers
Rank Reverse Distance Hazard Cumulative 20 19 17520 F 1/19 0.2385
rank (km) (1/rank) hazard 21 18 17540
1 38 6700 F 1/38 0.0263 22 17 17890
2 37 6950 23 16 18450
3 36 7820 24 15 18960
4 35 8790 25 14 18980
5 34 9120 F 1/34 0.0557 26 13 19410
6 33 9660 27 12 20100 F 1/12 0.3218
7 32 9820 28 11 20100
8 31 11310 29 10 20150
9 30 11690
30 9 20320
10 29 11850
31 8 20900 F 1/8 0.4468
11 28 11880
12 27 12140
32 7 22700 F 1/7 0.5896
13 26 12200 F 1/26 0.0942 33 6 23490
14 25 12870 34 5 26510 F 1/5 0.7896
15 24 13150 F 1/24 0.1359 35 4 27410
16 23 13330 36 3 27490 F 1/3 1.1229
17 22 13470 37 2 27890
18 21 14040 38 1 28100
19 20 14300 F 1/20 0.1859

35
 Example 1
Plot the data on log 2 cycle paper x log 2
cycle paper
Estimate Weibull shape parameter
Estimate Weibull scale parameter
Interpret results

36
 Example 1

Cumulative hazard plot for shock absorbers on linear

paper

30000
25000
distance

20000
15000
10000
5000
0
0 0.2 0.4 0.6 0.8 1 1.2
Cumulative hazard

37
 Example 1
Log cumulative hazard for shock absorbers

100000
log distance

10000

1000
0.01 0.1 1 10
log hazard

= 2.6,
= 28500km
R2 = 0.98
38
 Example 1
Since looking at one known failure mode use the estimated
parameters to fit to the distribution

Survival plot for vehicle shock absorbers with

Beta =2.6 and Eta=29000km
Probability of survival

1.2
1
0.8
0.6
0.4
0.2
0
0 5000 10000 15000 20000 25000 30000 35000 40000
kilometers

39
 Example 2:Hazard plot on linear paper

Cumulative hazard plot for O ring failures

2000

1500
hours

1000

500

0
0 1 2 3 4 5 6
cumulative hazard

40
 Example 2: Hazard plot on log paper

log cumulative hazard for O ring failures

10000 = 1.01
1000 = 360hrs
log hours

100 R2 = 0.98
10

1
0.01 0.1 1 10
log cumulative hazard

41
 Example 2: Interpretation of results
 = 1.01 is approximately an exponential
distribution and constant failure rate
 = 360 hrs = 1/ = Mean Time to Failure
Calculating the MTTF from the data gives:
– Total hours/number of failures
– 26839/73 = 367 hrs

42
 Example 2: Survival function
R(t) - Survival Function for O ring failures

1.2
Probability of survival

1
0.8
0.6
0.4
0.2
0
0 500 1000 1500 2000
Hours to failure

43
 Example 2: Failure Distribution

F(t) for O ring failures

1.2
Probability of failure

1
0.8
0.6
0.4
0.2
0
0 500 1000 1500 2000
Hours to failure

44
 Example 3 : pumps
pump no age at failure failure mode

Two dominant 1 1180 m

2 6320 m
failure modes 3 1030 i
Impeller failure (I) 4 120 m

Motor failure (m) 5 2800 i

6 970 i
7 2150 i
8 700 m
9 640 i
10 1600 i
11 520 m
12 1090 i

45
 Example 3: ignoring failure modes

log cumulative hazard for all failures

10000

1000
age

100

10

1
0.01 0.1 1 10
cum hazard

46
 Example 3: impeller failure

log cumulative hazard plot for impeller failures = 1.95

10000
= 1900hrs
R2 = 0.93
age at failure

1000

100

10

1
0.1 1 10
cumulative hazard

47
 Example 3: motor failure

log cumulative hazard plot for motor failures

= 0.76
10000
= 3647hrs
age at failure

1000

100
R2 = 0.978
10

1
0.01 0.1 1 10
cumulative hazard

48
 Advantages of Cum Hazard Plotting

It is much easier to calculate plotting

positions for multiply censored data using
cum hazard plotting techniques.
Linear graph paper can be used for
exponential data and log-log paper can be
used for Weibull data.

49
 Disadvantages of Cum Hazard Plotting
It is less intuitively clear just what is being plotted.
– Cum percent failed (i.e., probability plots) is meaningful
and the resulting straight-line fit can be used to read off
times when desired percents of the population will have
failed.
– Percent cumulative hazard increases beyond 100% and is
harder to interpret.
Normal cum hazard plotting techniques require
exact times of failure and running times.
With computer software for probability plotting, the
main advantage of cum hazard plotting goes away
50
 Summary
Important lifetime distributions
– Failure distribution (CDF), Survival function R(t)
and the hazard function h(t)
Some parametric distributions
– Exponential, Weibull and Normal
Weibull probability plotting
Distribution fitting using hazard plotting
techniques

51