Version 3 November 2005

VE School
Weibull analysis
Nicolas Forissier
january the26 th 2007
Version 3 November 2005
VE School
What is a Weibull analysis?
A Weibull analysis on warranty data consists in:
1 - Studying reliability as a function of mileage.
2 - Making the assumption that the Reliability law
follows a Weibull distribution
3 - Finding the Weibull distribution which is most
appropriate to the data observed.
Version 3 November 2005
VE School
Weibull distribution
is the most appropriate tool for reliability analyses
can be adjusted to most observed failure modes
is capable of describing each phase of the life cycle of a device by
means of a point-in-time failure rate given in its simplified form by:
The failure probability is:
The reliability at a given mileage is written as:
( )
1
|

\
|
=

t
t
What is a Weibull distribution?
|
|

\
|
|
|

\
|
=

t
t R exp ) (
|
|

\
|
|
|

\
|
=

t
t F exp 1 ) (
Version 3 November 2005
VE School
= scale parameter:
corresponds to the mileage at which there are 63.2%
of defective devices
is hard to interpret as it depends on the shape
parameter
= shape parameter:
characterises the phase of life
WEIBULL DISTRIBUTION: its parameters
Version 3 November 2005
VE School
wear, fatigue breakdown in youth
debugging
= 1 (or close to 1)
maturity
random
breakdown
useful life
< 1
youth
> 1
ageing
Characterising a product's phases of life
using the parameter
often process
faults
*Inappropriate process
*Non robust DT/process variability
*Product characteristic too dispersed
in relation to the functional requirement
often product design
flaws
If it appears prematurely:
*Design flaw
*Non robust DT/process variability

Version 3 November 2005
VE School
Weibull distribution: approximation of F(t) in 0
For a Weibull distribution: F(t) = 1- exp(-(t/)

)
development limited in 0 to the order 1 of F(t) is
F(t) (t/)

ie F(t) constant* t

therefore, if we multiply t by 2, F(t) is multiplied by 2

examples:
0,1%
0,1%
0,1%
0,1%
0,1%
F(10,000
km)
8 *0,1%=0,8% 3
4*0,1% = 0,4 % 2
2* *0,1% 0,283 % 1,5
2*0,1% = 0,2% 1
*0,1% 0,141 % 0,5
F(20,000 km) Beta
2
2
Version 3 November 2005
VE School
WEIBULL DISTRIBUTION: graphical
representation
Using the data observed and the calculation
of F, we change the variables:
Graphical representation:
( ) ( )
X t t = ln ( )
( )
Y t
F t
=

|
\

(
ln ln
1
1
and
F Y
X
t
The possibility of representing the
reliability law by a 2 parameter
Weibull model is judged according to
the "alignment of points":
NB: The Weibull straight line characterises just one failure
mode
( ) ( ) ( )

Ln X o d
t
t R t F
|
|

\
|
|

\
|
= = = Y ' , exp 1 1
( ) Ln X = Y
The parameter therefore corresponds to the gradient of the straight line
Version 3 November 2005
VE School
wear, fatigue breakdown in youth
debugging
= 1 (or close to 1)
maturity
random
breakdown
useful life
< 1
youth
> 1
ageing
often process
faults
*Inappropriate process
*Non robust DT/process variability
*Product characteristic too dispersed
in relation to the functional requirement
often product design
flaws
If it appears prematurely:
*Design flaw
*Non robust DT/process variability
F F F
Characterising a product's phases of life
using the parameter
Version 3 November 2005
VE School
Description of Weibull models
The Weibull distribution is represented
graphically according to different models
3 types of models encountered 3 types of models encountered
1 1 Mode
2 2 Parameters
1 1 Mode
3 3 Parameters
2 2 Modes
4 4 Parameters
Version 3 November 2005
VE School
Distribution at 1 mode 2 parameters Distribution at 1 mode 2 parameters
Representation of the 2 parameter 2 parameter Weibull
distribution describing one failure mode one failure mode
F t
t
( ) exp =
|
\

|
|
\

|
|
1

Description of Weibull models

Version 3 November 2005
VE School
Example of a line, for 1 mode 2 parameters with a Beta = 0.57 (therefore < 1).
The Weibull curve below shows the failures of the Turbo GARRET GT 15-44 mounted on engines X,
delivered during the 11/99-10/00 period. These failures are manifested in the engine's lack of power.
Example of a line
C
u
m
u
l
a
t
e
d
p
e
r
c
e
n
t
a
g
e
o
f
d
e
f
e
c
t
i
v
e
d
e
v
i
c
e
s
Multiplier coefficient : 100 km
Weibull curve
Description of the analisis :
Vehicle : all vehicles
Engine : engine X
Delivery period : No 99 Oct 00
NITG : M430 M431 turbo compressor
RC : shortage of power
Country : France
Version 3 November 2005
VE School
Example of a line
Analysis of this case showed that the
problem is due to the process.
Poor placement of the silicone seal
between the crankshaft and the casing.
we find silicone particles in the oil.
the turbo lubrication drilled holes are
blocked
the turbo is blocked
the customer notices a shortage of
power
Cylinder casing Crankshaft
bearing 1
SILICONE DEPOSIT
Silicone seam
When the bearing 1 is tightened to the crankshaft, the
siliconeoverflows into the cylinder casing
Version 3 November 2005
VE School
Description of Weibull models
Distribution Distribution at at 1 1 Mode Mode 3 3 parameters parameters
Representation of the 3 3 parameter parameter Weibull distribution describing
one one failure failure mode mode
|
|

\
|
|
|

\
|

=

t
t F exp 1 ) (
= positioning parameter,
corresponds to the mileage from
which it is considered possible that
there be a failure.
This is usually set at 0.
Taille du parc : 60226 Mode de calcul : 1 mode/ 3 paramtres
Nombre de points : 31 Paramtre de forme (beta) : 0,98
Kilomtrage max. des incidents : 80834 Dure de vie caractristique (eta ) : 47198873
Paramtre de position (gamma) : 6448
Coefficient d'ajustement : 0,985
B(0,5) en km : 223725
B(0,03) en km : 18879
Nombre de points hors cadre : 0
1000 5000 10000 50000 100000
0.0001
0.0002
0.0005
0.001
0.002
0.005
0.01
0.02
0.03
0.05
0.1
0.2
0.5
0.7
1
Km
%
c
u
m
u
l
d
e
d

fa
illa
n
ts
Version 3 November 2005
VE School
Distribution at 2 modes 4 parameters Distribution at 2 modes 4 parameters
Representation of the 2 2 parameter parameter Weibull distribution
describing 2 failure modes 2 failure modes
Mode 1
distribution
Mode 2
distribution
Note: the model used in West is different
West Model:
F( t ) = 1 R1(t) * R2 (t)
with R1(t) = 1 F1(t)
with R2(t) = 1 F2(t)
Description of Weibull models
Cm for mode change
t t if

t
exp 1 F(t)
t t if

t
exp 1 F(t)
Cm
Cm
2
2
1
1
>
|
|
|

\
|
|
|

\
|
=
<
|
|
|

\
|
|
|

\
|
=
Version 3 November 2005
VE School
Proportion of defective devices -
WEIBULL
Recovery of the
mileages at which
incidents occurred
from the warranty
database
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0
5000 10000 15000 20000 25000 30000 35000 40000 45000 50000 55000 60000 65000 70000 75000 80000 85000 90000 95000 100000 105000 110000 120000 125000 130000
mileage distribution
Identification
of the vehicle's
mileage profile
Modelling using the
WEIBULL distribution
F: Proportion of defective
devices
as a function of mileage
Estimate of the proportion of defective devices among
customers
Cleaning up
the data
Recovery of the
sales volumes
from the warranty
database
Calculation of the surviving devices
Estimate of the failure rates
at incident mileages
Incident milage
...
15 264
15 268
15 271
...
...
...
103 305
110 212
120 114
136 783
144 230
...
Version 3 November 2005
VE School
Weibull modeling: Uses and limits
Uses Limits
* Studying a proportion of defective devices as a function
of mileage
* Estimating a proportion of defective devices at a target
mileage.
* Identifying multiple failure modes
* Pointing to the sources of failure modes (product,
process faults)
*Specifying the mileage from which the element begins to
fail
* Extrapolating at high mileage!

And consequently:
- Identifying extended parts, for a new project, which are
not on target.
- Comparing, for an element or group of parts, the
reliability for different technologies.
- Assessing the effectiveness of counter-measures.
- Estimating, from the first incidents, future guarantee
costs.


Weibull analyses must be done on carefully selected,
uniform domains

The value of Beta is an indication as to the source of
the failure: it does not give us the failure mode..

Do not be content with values taken from the Weibull
curve, carry out technical investigations with the core
competences in order to back them up!!


Extrapolations at high mileage are highly uncertain
(due to the low number of incidents and the limited
known range).
The extrapolation may underestimate (or
overestimate) the proportion of failures.



Version 3 November 2005
VE School
back up
Version 3 November 2005
VE School
Maturity
(random breakdowns)
Youth, debugging
Age, wear
: Point-in-time failure rate
t (km, cycles, h, ...)
Point-in-time failure rate
The conditional failure probability between t and t+dt, knowing that the
system functions up to mileage t.
It is estimated
from the data:
Number of defective devices
between t
j
and t
j+1.
Number of surviving devices at t
j
.
(t
j
) =
n(t
j
)
Ns(t
j
)
The failure rate characterises the phases of product life
Version 3 November 2005
VE School
t is the physical factor for which we are studying the reliabili t is the physical factor for which we are studying the reliability of the ty of the
system system
R(t R(t)= )= Prob Prob ( (t t
fail fail
>t) >t) F(t F(t)= )=Prob(t Prob(t
fail fail
<t) <t)
R R and and F F characterise characterise the the life span life span distribution for a device ( distribution for a device (component, component,
system) system) until until the first failure under the first failure under given conditions given conditions
of use of use
mileage, number of
cycles
F
estimate
d
10%
Estimation of the proportion of defective devices at a given mileage: F
estimated
Estimating the failure probability
Failure probability law Failure probability law
Version 3 November 2005
VE School
Algorithm: Mileage profiles
They have been established by the DQGR from incidents observed in
previous years, mileages recorded during Red OTS or technical
inspections, mileages declared by customers during surveys...
They are defined by geographical area, energy (petrol or diesel) and
segment.
They are modelled by log-normal distributions.
Example of the Lagunas in France at 1 year of running
Petrol:
Average = 18,000 km
Standard deviation = 11,000 km
Diesel:
Average = 30,000 km
Standard deviation = 20,000 km
The mileage profiles at n months are deducted from the rolling
distributions at 1 year by rules of 3.
0%
20%
40%
60%
80%
100%
0 20000 40000 60000 80000 100000
Km
%

d
e

v

h
i
c
u
l
e
s

q
u
i

o
n
t

r
o
u
l

p
l
u
s

d
e

X

k
m
M2 Diesel 1 an
M2 essence 1 an