Determining the shape parameter of a Weibull distribution from

mechanical damage models

Fabrice Guerin ISTIA .Angers

0

Bernard Dumon ISTIA Angers

Ridha Hambli ISTIA .Angers

Key Words : Weibull distribution, Log-normal distribution, Mechanical damage, Shape parameter, L-imitedfatigue.

The stage I11 is characterized b y different mechanical dam-

SUMMARY & CONCLUSIONS
ages like fatigue, wear, tear or corrosion. In literature, links are
found between the values P and the damage modes (ref 8 ) .
The aim of this paper is to show that it is possible to de-
termine the shape factor of the weibull distribution relating to P E [ 1.5 ; 2.51 The damage phenomenon looks like fatigue
a mechanical damage model. In particular, we defiie the shape P E [3 ; 41 The damage phenomenon looks like wear and
parameter linked to damage by limited fatigue. Also, the pa- tear or corrosion
rameters of distribution of cycles number to failure is charac- In the case of particular mechanical components, we have
terized using the limited fatigue approach by Basquin model. a value of P
After, the Weibull parameters (P and q) are estimated by the Ball bearing : P = 1.5 (ref 7, 10)
method of moments. This work is currently used for other Gear : P E [0.8 ; 1.3 3 (re1'8, 9)
modes in order subsequently to obtain more precise conclu- However, all these results are based on experience. In this
sions when one analyses a Weibull sheet. article, we therefore suggest to justify the value of P in the
case of mechanical damage by theory by using mechanical
damage models as a base.
1. INTRODUCTION
Notation list
The distribution of the time to failure is commonly defined
by a Weibull distribution characterized by : P the shape parameter of weibull distribution
(,$ rl the scale parameter of weibull distribution
N number of cycles to failure
P (-
f ( t )= - t)o-'J-J t time to failure
r\r\ P mean
P the shape parameter
with : 0 standard deviaf ion
q the scale parameter PN mean of the ratidom variable N
The Weibull distribution defines correctly the different ON standard devialion of the random
stages of component life (see Figure 1). variable N
PLO~(N) mean of the random variable Log(N)
(J standard deviaiion of the random
a,
*
m ! variable Log(N)
1 3 1 \ I
Rn, yield tensile stress
(JD endurance limLt
k and A parameters of I3asquin model
(Jm mean stress
(Ja amplitude of s tress
a the probability of breaking
k(a) and A(a) Basquin parameters for a a given
t
N(a) number of cycles for a a given
Figure 1 : Mechanical component life cycle. amplitude of stress for a a given
o a ( 4

R.L. Doyle (ref 13) explains that the majority of a me-

chanical component's life spent in stages I1 and 111, which
show a gradual increase in failure rate during stage I1 and an
accelerated failure rate during stage 111. This is brought about
by the fact that mechanical component have continuous wear
throughout their operating life.
0-7803-6615-8/01/$10.000 2001 IEEE
156 2001 PROCEEDINGS Annual RELIABILITY and MAilNTAINABILITY Symposium
 2. DEFINITION OF THE SHAPE PARAMETER FROM
THE MODEL OF MECHANICAL DAMAGE.

For several damage models (fatigue, crack, creep, wear and

tear), the time to failure is generally characterized by a Log-
normal or Weibull distributions. The following examples (ref
2) present the analysis of time to failure by Weibull distribu-
tion (see Figures 2.a and 3.a) and Log-normal distribution (see
Figures 2.b and 3.b) for a gear and ball bearing.
0.1 I I I I I I I l l
2 3 5 6 7891MM
'O0. I

zot Time to failure

Figure 3.b : Log-normal distribution of time to failure of
ball bearing
These examples show the possibility to define the time to
failure by Weibull or Log-normal distributions.
The interest of Weibull distribution is to define all stages
of component life (see Figure 1). But the difficulty is to link to
a p value the damage mode. So, we propose to evaluate the
0.2 Weibull parameters in considering a mechanical damage, in
I I I I I I I I I
Im0.1
2 3 L s 87asmam particular the fatigue (see $2.1). In this case the time to failure
is defined by a Log-normal distribution (ref 2, 3, 4, 5). Then,
Figure 2.a : Weibull distribution of time to failure of gear we propose to evaluate the Weibull parameters by the method
of moments (see $2.2).
2ot
2.1 Implementation to fatigue
As part of this paper, we suggest to clarify the approach by
studying the fatigue (ref 2, 3, 5, 11) which appears when the
component is put under cyclic-stress loading (see Figure 4).
The cycle is characterized by :
G,,,the mean of stress
G, the amplitude of stress

Time to kllhrre
Figure 2.b : Log-normal distribution of time to failure of gear

20
3a t t-
Figure 4 : cyclic-stress loading

The design of this type of component takes place whilst

considering the endurance limit G~ characterized by a Wohler
curve (see Figure 5). This curve allows us to define the am-
plitude of the life cycle not to overtake to avoid breaking be-
fore a given time. The fatigue is characterized by 3 areas :
oligocyclic fatigue, limited fatigue and unlimited fatigue.
2
0O 1m
2 3 4 5 6 783lwO
Figure 3.a : Weibull distribution of time to failure of ball
However? this fatigue phenomenon is distorted by a certain
dispersion and it is common to have breaking equiprobability
bearing curves appear with the Wohler curve. It is thus admitted that
for a given stress the number N of cycles to failure follows a
Log-normal distribution (or the variable Log(N) follows a
normal distribution) in the area of limited fatigue.

2001 PROCEEDINGS Annual RELIABILITY and MAINTAINABILITY Symposium 157

 2.2 Determining the weibull parameters by the method of 2.3. determining the shape factor in the case of limited fa-
moments tigue
The method of moments (ref 1) is one method of estimat- The Log-normal distribution of N is characterized by the
ing distribution parameters using the moments of samples. The parameters p L o s ~and
) cLogV).From this data we can infer the
first moment (mean) and second moment (variance) for mean and variance of N by :
Weibull distribution are defined by :

p =qr(i++)

The Weibull parameters are evaluated in solving these In paragraph 2.2, we have determined ;he Weibull pa-
equations with two unknowns (0 and q). The coefficient of rameters by the method of moments. The shape factor is de-
variation (6) based on equations (4) and ( 5 ) is : fined by the equation (13) with x equal to :

This equation however does not allow us to determine

analytically the value of p. To this end, we have defined dif-
ferent values of variation coefficient 6 and searched by
mathematical optimization the values of p justifying the equa- 4. EXAMPLE
tion (6). Finally, we obtain one graph'(see Figure 9) which
gives the value of p depending upon the coefficient of varia- We will study an aluminum component (2017) with the
tion. mechanical characteristics defined in Table 2.

M;:an Standard de-

viation
Result of optimization Yield tensile stress 387.MPa 20 MPa
Endurance limit 122 MPa 10' cycles 10 MPa
2.5

_________
The typical points (see Table 1) of the straight lines of
Basquin (for a=16%, SO% and 84%) are characterized in Ta-
ble 3.
0.5 1- I Stress I
0 i
0.97 1.17 1.37 IS7 1,77 1,97 2.17

6 *+1
Figure 9 : Graph which gives p depending upon the coefficient
of variation 84%

We have then searched a link which describes this curve and

this lead us to :

' 2.4
= 8.Sx - 2 . 2 -~6 ~
with ~ = 6 ~ + 1
The Basquin parameters (defiined by relations ( 6 ) and (7))
for the 3 straight lines are summarized in Table 4.

When we know 0 and p we deduce the scale parameter q

by :
ci = 50%
(14) ci = 84%

Table 4 : Basq-uinparameters

158 2001 PROCEEDINGS Annual RELlABlLlTY and MPJNTAINABILITY Symposium

 Table 5 : Different values of computing

Then we can define the values of the shape factor p for the 3 . Kececioglu D., Jiang M. and Sun F-B.., G A unified approach to ran-
dom fatigue reliability quantification under random loading )), Proc IEEE
different values of stress. The results are presented in Table 5 Reliability and Maintenability, 1998
and illustrated in Figure 9. 4. W. Nelson, Accelerated Testing, 1990 ; John Wiley & Sons.
5. R.E. Little and J.C. Ekvall, Statistical Analysis of Fatigue Data ; 1979 ;
P=f(o) American Society for Testing and Materials, STP744.
6. RADC T64-50 : Reliability Predictions for Mechanical and Electrome-
chanical parts.
7. NF E 22.394, French norm
8. Robert B. Abemethy, The New Weibull Handbook, ed. Gulf Publishing
Company, july 1996
9. E.J. Wellauer, How long will your gears last, July 1963, Plant engi-
neering.
IO. Harris, Rolling Bearing Analysis, ed John Wiley and Sons
1 1. Shigley, Mechanical Engineering Design, ed Mc GRAW Hill
12. MIL HDBK 5 B
13. R.L. Doyle, "Mecanical Reliability", RAMS Tutorial Notes, 1991

100 150 200 250 300 BIOGRAPHIES .

(r

Figure 9 : Evolution of p relating to loading stress Fabrice GUERIN, Assistant Professor,

1.S.T.I.A - Quality and Reliability department
We thus notice the evolution of the shape parameter p 62 av Notre-Dame du Lac
49000 Angers-France
linked to the stress of loading. Indeed, the lighter the loading
the more p decreases, which is explained by the fact that when Intemet (e-mail): [email protected]
one loads a lot, one damages quicker (p will be stronger) and
Fabrice GUERIN is an Assistant Professor at University of Angers, he
when on the contrary one loads less, the less one damages (p teaches a course on Mechanical technology, Statistic, Metrology, Reliability,
will be weaker). When the loading stress decreases, the p ... His interest lies mainly in mechanical reliability, bayesian statistic and
value leads to 1. The mechanical damage observed is not the accelerated testing.
limited fatigue but the unlimited fatigue (corresponding to
Bemard DUMON, Professor,
random failure). Member of IEEE and ASQC
1.S.T.I.A - Quality and Reliability department
62 av Notre-Dame du Lac
5 . CONCLUSION 49000 Angers-France
Intemet (e-mail): [email protected]
This article shows that it is possible to define the shape pa-
Bernard DUMON is a Professor at University of Angers, he teaches a
rameter of the Weibull distribution relating to a mechanical course on Reliability, Markov chain, Bayesian Statistic, .... His interest lies
damage mode (limited fatigue). This work is currently used for mainly in reliability, bayesian statistic and accelerated testing.
other modes in order subsequently to obtain more precise
Ridha HAMBLI, Assistant Professor,
conclusions when one analyses a Weibull sheet. It will thus be
!.S.T.l.A - Quality and Reliability department
possible to gain a more precise idea of the damage modes of a 62 av Notre-Dame du Lac
product without any physical analysis of the faulty products. 49000 Angers-France
lntemet (e-mail): [email protected]

REFERENCES Ridha HAMBLI is an Assistant Professor at University of Angers, he

teaches a course on Manufacturing Technology, Science of material, CAD, ....
His interest lies mainly in reliability and numeric simulation (FE)
1. B.M. Ayyub and R.H. - McCuen, Probability, Statistics & Reliability
for engineers, CRC Press New York, 1997
2. J.H. Bompas-Smith, Mechanical Survival : the use of reliability data,
Mc Graw Hill book, 1973

2001 PROCEEDINGS Annual RELIABILITY and MAINTAINABILITY Symposium 159

 From this model we can infer the characteristics of the log-
Oligocyclic
aa & fatigue.1,
limited fatigue , unlimited fatigue
r,.
normal distribution of the number of cycles to failure. In this
-!- way, one can infer the parameters of the Basquin model from
Distribution of lifetime
the two points of the curve (see Fizure 7).
Distribution proportional

for 50% breaking

I ,
/ Distribution of aD
I

Log N
Figure 5 : Wohler curve Lo@ I ) Lorn,) LOkW
Figure 7 : Simplified building of the Wohler curve
This curve can be defined by different models (ref 5) in-
cluding the Basquin one (valid in the area of limited fatigue).
The Basquin model is defined by :
(4)
~ J N = A (2)
with : A constant
oaamplitude, constant value of the stress (5)
k factor linked to the material After this, it is possible to tackle the probabilistic angle by
N number of cycles studying the normal laws associated to the yield tensile stress
This curve is represented in logarithm scale by a straight and endurance limit. Thus one can determine new typical
line. The equation of this straight line is : points (oal(a),NI,q 2 ( a )and N2) for a given breaking risk (a)
1 1 and infer the parameters of the model by:
Log(0 a ) = - - Log(N) + - Log(A) (3)
k k
Amongst literature (ref 1l), one finds simplified building
method of the Wohler curve from the Basquin model. Thus the
details of two typical points are indicated for steel and alumi- (7)
num in Table 1.
Thanks to this data it is possible to define the breaking
equiprobability curves for different values of a.They enable
For steel For N=10 we have ForN=10 wehave
us, under a given stress, to obtain the time to failure for each
0.95R,1,
value of a.
For aluminium For N=10 we have For N=10 we have
0.8RllI
Table 1 : Cha ,acteristic points of Wohler curve (Handbook
information (ref 11, 12))

In the case of a material (steel or aluminum) and

when we know its yield tensile stress R,,, and endurance limit
OD, we can build the simplified straight line with breaking
equiprobability curves (see Figure 6) knowing the dispersions
on R,,, and OD. I
Distribution proportional
Figure 8 : definition of the log-normal distribution of the
Distribution of lifetime number of cyclcs to failure
under constant stress

Distribution of an Thus when one determines number of cycles to failure for

a equal to 16%, 50%-and 84% (50% equals to pLosp),16% to
OL,,~(N)
~ L ~ ~ ( N ) - and 84% to p L O g : ~ ) + one defies the
log-normal distribution (see Figwe 8) :
P L o g ( N ) = Log"O%)) (8)
log(N)
Figure 6 : Basquin model
Log(N) = (
Log(N(84%)) - Log(N(16%))
2
) (9)

160 2001 PROCEEDINGS A n n u d RELIABILITY and MA.INTAINABILITY Symposium