Abcdefg Adds

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

e Pergamon

Micr()("Mr/ru", Re/iab .• Vol. 37. No.6, pp . 879-883, 1997

PIT: S0026-27 14(96)00253-3


:D 1997 ElseviCT Sci.nee Ltd
Printed in G1"<'at Britain. All rights reserved
002S-27 14j97 $17.00 + 000

BAYES IA N T ECHNIQ UES TO RED UCE THE SAMPLE SIZE


IN AUTOMOTIVE ELECTRO NICS ATTRIBUTE TESTING
ANDRE KLEYNE R,t SHRI KAR BHAGATH,t MAU RO GASPARI N Lt
JEFFREY ROBI NSON§ and MARK BENDERt
t Delco Elc.:lronics Corporation. Kokomo , IN 46904-9005. U,S.A.; t Deparlmenl of Statistics, Purdue
University. West Lafayette. IN 47907- 1399, U.s,A. , §Mathematics Department, General Mo tors Researc h
LabOratories, Wa rren. M[ 4&090-9055, U.S.A.

( Ref/'h'ed/or pllbiic(lIion 2 July 19%)

Abslract- This paper discusses the application of Bayesian techniques 10 the determination of sa mple sizes required
for an attribute lest of a product in order to demonstrate a target reliability with a spedficd confidence. The mcthod
is lxlsed on analyzing statistical data on similar prodocts and incorporating them into a Bayesian prior distribution
for the unknown reliability. A mixture prior obtained by combining a bela prior with the uniform rectangular prior
(representing the unkno"'n content of the new product design) is discussed. The suggested method can significantly
lower sample sizes for attribute It"Sts and thus reduce cost, time, and resources currently being spent on reliability
demonstration testing. A numerical example at the end uf the paper illustrates the method. to 1997 Eh;cvier Science
Ltd. All rights reserved.

INT ROl) ucn ON specified mission life under standard conditions: an


attribute reliability experiment is performed to learn
In the pursuit of high quality and high reliability in more about it. The experiment consists of observing
a mass production Cllvironment, the au tomotive N successes out of N reliability test trials. A peculiar
manufacturers require their suppliers to prove a feature is that most often no less than a 100% success
target reliability with an assigned confidence level on rate is required- failing which corrective actions are
a supplied product. This is usually done through a to be taken- whereas in the usual reliability trials the
reliability demonstration test by running a certain success ra te, albeit usually high, is random.
number of samples under conditions simulating the Techniques commonly utilized to calculate spmple
mission life, an experiment which is sometimes called sizes for a reliability demonstration of a product
lest to a bogey. Most of the time the sample size is when a 100% success rate is required are generally
determined only by the requi red reliability and the refe rred to as Success Run fo rmulae 11,21. The
confidence level. Most of the methods currently used likelihood function, that is the probability of
in the industry prcsume no prior information about observing all successes given a certain value of the
the product or its predecessors. though very often this unknown product reliability R, is
information is available. With the ever increasing
reliability requirements the sample size to be tested is L(dataIR) = R". (1 )

growing out of proportion and economical sense, Based on this eq uation. the expression
requiring larger and larger amounts of human
resources and capital equipment. Based on the fact C= I - R ~ (2)
that ma ny new automotive products are evolutionary is known as the "Success Run formula" [2] and can
and not revolutionary, Bayes method can be onc of be obtained from the elassical Clopper and Pearson
the approaches to incorporate prior knowledge about (3J approach to the determination of confidence
the product, thus reducing the numbe r of test samples intervals for a binomial proportion. In eqn (2), RL
and the amount of resources dedicated to the test is the lower bound of a one-sided ex 100%
programs. confidence interval for the unknown reliability R.
Rl is referred to from now on as the demonstrated
EXISTING TEC HN IQ UES )-' OR SAMPLE SIZE reliability. In the automotive indust ry C and Rl
DET ER1\'II N A"J"IO N
are usually stipulated by the customer, the Success
Statistical experi ments are gencrally performed Run formula is then used for the dete rmination of
to learn mo re about unknown parameters charac- the required sample size N.
terizing our material of interest. In an automotive Using a Bayesian approach instead, we use
setup, the unknown parameter is the product prior distributions on the unknown parameters of
reliability R, that is, the probability of surviving a a statistical experiment to exploit useful pre-

879
880 A. Kleyner {'I 01.

experimental information, e.g. the d ata from binomial likeli hood such as eqn (I), a beta prior
previous test resul ts or similar product usage. For distribution on R, wi th density
Success RUll ex perimen ts, the likelihood (I) has to
W ·'(l - R)' · '
be combined with the prior distribution on R to
obtain a posterior dist ri bution on R. Such a poS+
/I*(R) = PeA. B) if 0 *" R *" I (4)

(crior distribution summa rizes all availab le infor- where


mation about the unknown product reliability R.
T he Bayesian version of the classical Success P(A B) ~ f(A )flB )
, f(A + B)
R un formu la uses a uniform prio r, a lso called a
Rectangular Prior, which presumes an equa l likeli- is particularly convenient; the constants A and B
hood for the reliabUity value to fall anyw here (sometimes called hypcrpa rameters) have a nice
between 0 and I, and expresses the idea of " vague"' interpretation- A being tho ught, sometimes, as the
prior mformation. In other words, since this prior number of successes out of A + B trials in a similar
assigns the same weight to every val ue of R, we p re-experiment , real or imaginary. Morc important ly,
expect it to produce results similar to the classical the beta p rior distribution is conjugate to binomial
Success R un formula. sampling, that is, the postcrior is a beta distribution
Combini ng the unirorm prior and the likelihood, a s well. T his allo ws for a continuous updating of the
using Bayes theorem, we o btain the Bayesian ve rsion posterior within the samc general class of d istri-
or the Success Ru n formula from the posterior butions. T he uni form prio r is a special case of eqn (4)
probability for A = B = 1.
The posteri or density 'Jfl R o btained by combining
eqns (I) and (4) through Bayes theorem is
' R" dR
C = p(RL< R <
J.
I) = " I - Rt'· ' (3) R"R'< · '( I - R)'. ')
( P(A. B)
1' R" dR n*(Rldata) = --.:-~~---..:.-
fS"'S' · '(I - S)8. l dS)
( P(A . B)
whe re C is t he coefficient of the cred ible interval
[Rl. I]_ Rl • the ( I - C) quantile of the posterior RHN · '(I R)-· I
distribu tion of R, is still to be referred to as the (5)
(Bayesia n) demonstrated reliability. An interpret-
{J(A + N , B)
ation of eqn (3) is that, after the successful that is, a beta density with parameters (A + N) and
completio n of a Success Run experiment with N B. T he usc o f bcla prio rs for binomial sampling has
uni ts, the unknown reliabi lity R lies in the interval a long history, starting somewhere in the prehistory
[RI.. I], with C x 100% probability. o f modern Bayesian statistics. For an acco un t of the
The sampk size ca lculated using eqn (2) is one uses of beta d istri butions in attribute reliability tria ls,
sample mOfe than what we would get using eqn (3)_ see Martz and Waller [4, 5]. As in thc case of the
Some comlllon reliability demonstration req ui re- standard Success Run formu la (3), the immediate use
ments a nd the sample sizes for the Success Run of of posterior (5) is to establish a reliability level RL
these demo nstrations ,n e given in Table 1. Equation above which there is a high Bayesia n credibility C
(3) has IX."'t!n used in the calculation of these sample that the reliability R will be met. For this purpose, we
siles. use equation

c ~ J.'
RH N· '( I _ R)'·'
fRO !\1 RETA PRIORS TO MIXTU RES OF BETA (6)
PRIORS ~'OR PRQIJ UCr RELIABILITY {J(A + N , B) dR
"
A generali za tio n o f the Success Run fonnula (3) which tells us that there is a C posterior probability
can be obtained from priors other than the uniform. that R wi ll be grealer than R l ,
In Bayesian statistics, it is well known that for a If. before Ihc experiment. we require a certai n
Bayesian credibility C based on the contract ual
Table I. Some common rf:liabitity demonstration specifications, for a given A and B the o nly unknown
requirements in ex pression (6) is thc sample size N. For a given
Sample size prior (4) we have to solve eq n (6) n umerically fo r N ,
Reliability to be Confidence (Success Run in order to know how large a sample size we have to
demonstrated level formula) observe, with a 100% success rate, to satisfy the
0.95 0.9 45 required C and RL •
0.97 0.7 40 The choice of the parameters of the prior A a nd B
0.99 0.5 69 is a erucial one. [\ seems reasona ble, in automo ti ve
0.99 0.9 229 reliabili ty applica tions, to base such a choice on
Automotive auribme testing 881

failure data, which a re easily available and contain a Notice that the usc of a uniform prior a lone would
lot of relevant information on past mooeh or similar lead to the Bayesian version of the Success Run
products. In the presence of information on the formula ; the use of mi~tures therefo re represents
success rate of n previous life tests, a possible way to a reasonable compromise between Bayesian and
obtain A and B is based on an empirical Bayes classical mcthods.
approach discussed in Ref. [61. See Martz and Waller The idea of using mixture priors in the conte~t of
[41 whcrc empirical Bayes estimates of A and B arc product reliability could be genenllized to the case
derived (Appendix). of heterogeneous prior information, in particula r to
Beta priors of the form (4) have a long history and thc case where failure data are available for differ-
arc mathematicall y convenient, but for our purposes ent past products, s()me more similar than others to
they arc too restrictive. T he best way to understand the new product. In that case, the analysis could be
this is by observing that an industrial product is in generali zed to the consideration of prior densities of
continuous evolution and, although a lot of similarity the form
exists between old and new Inodels, we always have
a margin of novelty which should be accounted for.
RA,- ' X (1 R)I. - ')
On the o ther haod, we do waot to use prior rr(R) = ~ ( p, P(A" B,) + (I - p) (8)
information on similar products in our research; this
is the reason why we want to use Ba yesian methods
in the first place. The right compr()mise between these where P = l:, PI and the different knowledge co-
confl icting goals seems to be 10 generalize the class of efficicnts Pi reflect d iflerent degrees of similarity
beta priors to the larger class offinite mix tures of beta between the new and old products. Another reference
priors. The plan is then to put wgelher a prior to the use of mixture priors in Bayesian reliabilit y is
distribution derived from failure data and a margin Savchuk and Martz 17J.
of uncertainty intrinsic to the new model. The laller Only mixtures with two components are con-
margin of uncertainty can be e~presscd as a uniform sidered here. Combining eqns (I) and (7) using Bayes
prio r on the reliability. theorem we o btain the posterior density,
Our proposal is therefore to use a two-component
mi~ture of beta distributions, with density

TI(R) = P
RA - '
x (I
fI(A, 0)
R)--'
n(Rldata) = ~~:!.~;'~.+~~R~·~·~·r'-~'+~X~(I~~R=)='=-:'
(I p)R
(I p)
p
P(A
P(A,O)
N, B) (9)
N+ I + p P(A,B)
+( l -p) if O ~ R ~ 1. (7)
and the corresponding e~pression
The first componenl of the mixture is a beta prior
with parameters A a nd 0 to be derived from failure
data. The second component of the mixture is a
uniform prior (a special case oflhe beta) representing
c= r' n(Rldata) dR (10)

uncertainty about the new product reliability. The


j"
two components arc combined acco rding to weights
p a nd (1 - p), where p is a ·'knowledge factor'· where a required demonstrated reliability Rl and
representing how similar the new product is to the o ld credibility C can be achieved. The solution of
one, and (I - p) is an "innovation faclor" , reflecting eqn (10) has to be found, in general, by numerical
the proportion of new content in the new product. methods.

Table 2. Calculation of eoeflicients A and B from IPTVJreliability data


IPTV Volume sold I, Reliability Rj
Model year Body style (mission life) (x 1000) 1- (iPTVfl OOO)
Product
Type I 2.94 170 0.9971
Product A 19XX
Type ][ 3.57 121 0 .9964
2 0.9976
Type III 2.45 206
3 0.9941
4 Type IV 5.32 l5
0.9976
5
6
7 19YY
Type V
Type VI
Tyoc I
2.38
8.68
1.75
38
306
" 0.9913
0.9983
Type II 1.12 1]3 0.9989
8 0.9959
9 Type III 4.06 87
Type IV 1.61 27 0.9984
10 0.9989
I] Type V 1.12 173
Type VI 4.4\ 156 0 .9956
12 K .. rl,~ I .. \.66£ - 04 rR,_ 11.96
n = 12
A ... 769.34; B = 2.53.
882 A. Kleyner r:t al.

1
r- - - ~ - -r- - -
- Uniform
0.8 f
Lll
~o .
u
6

0 .4
-
Mixture 150 Beta/50%Unifor
- - -
)
'- - - /
0.2 II
o
Beta /
0.98 0 .9 85 0.99 0.995 1

Reliability R
Fig. I. COFs for beta. mixture and uniform d istributions with A _ 769.34 and B ", 2.53.

AN t;XAMPLE TO DEl\'1ONSTRATE A PPLICA TION fac tor p = 0.1), the sample size reduces to 54 and as
OF THE TECIINIQ Ut:
the knowledge factor increases, the sample size
The sample size determination technique descri bed docreilscs.
in previous sections of this paper has been ap plied
10 a re,11 life example in order to demonstrate a
CONCLUSION
significant reduction in sample size. Table 2 shows
failure data for an electronic vehicle control product The method prcsented in this paper ,has great
(slightly modified frOID actual data for security potential for cost reduction in reliability demon-
reasons) in terms of [pry ( Incidents Pcr Thousand stration testing in a mass production environment,
Vchicll."S). Table 2 shows breakdown by model years like a n automotive electronics industry. The fai lure
and body styles, totally const ituting 12 test sets data on similar products used to build a "prior" can
(/I := 12). The observed fililure rates R;, are calculated significantly dec rease the number of test items to a
from the IPTV data usi ng: bogey. Even in cases with a low knowledge factor
such as 0.2 or 0.3 (20-30% prior knowledge about
(1J ) the product), the method may present significant
sample size reductions.
Using equations from the Appendix, the values of A [n cases with a favorable prior, the sample size
and B for the data in Table 2 a re found (0 be 769.34 may someti mes go down to zero or even become
and 2.53, respectively. The cumulative distribution
functions (CDFs) of the uniform, beta ilnd mixture Tab[e 3. S3mp[e si~es for v3riOllS knowledge
distributions are shown in Fig. I for the crucial range faclOrs al R "" 0.99 ~nd C .. 90%
of 0.98 ~ R~ I. Knowledge factor (p) Sample size N
Using eqn (10) and solving numerically for the
sample size, N, for a demonstrated reliability of
I o
0.9 1
RL "" 0.99 with C = 90%, the sample sizes for various 08 2
knowledge faclors arc as shown in Table 3. Using the 0.7 4
classical Success Run formulil (no prior knowledge 0.6 6
about the product or k.nowledge factor p = 0), 229
0.5 9
0.4 \3
samples of the product A will have to be tested with OJ 19
no fai lures to demonstrate a 0.99 reliability with 90% 0.2 30
confidence. From Table 3 it is seen that with only a 0 .1 54
10% prior knowledge of the product (knowledge o 229
Automotiv~ attribute testing 883

using mu ltiple ,ources of prior information; binomial


negative. The zero o r negative sample sizes would
$lImpling. IEEE Trnnsac/ions 011 Reliability. Mareh
mean that the requi red reliability has already been 1994, 43, 138- 144.
demonstrated during the previous stages of product
development and no furt her testing is needed. APPEl\O IX
In instances with an unfavorable prior the num~
Martz and Waller [4} derive empirical Bayes estimates of
ber of s;lmples to be tested may actually exceed the
A and B as foll ows:
num ber computed using the elassical method . This
means that the product's prior has already shown
that the product 's reliability is most li kely less
than the desired o utco me a nd no further tes ting (AI)
should be perfomled wi thout ap propriate design
corrections.

Acknowledgements- W e would like to gratefully acknowl- and


edge the help and support w~ rCl;:eivcd from Joc Boyle. A = (A + B)R
Thomas Torri, Jay Rosen and Ted DeGarmo at Delco
Electronics; and Joe Wolkan at the Gen~ ral MOlOrs Proving whe~e n is the n umber of life tests; Ii is the number of units
Grounds. in the i' test: R, is the r
observed failure rate

N umber of fai lu res in the r test. .


K = l:l,- I
REF ERENCES = Ij ,. ,
1. Johnson. L. G .• G MR ReI. Manual, GM R-302, and
Gencral Motors Rcsearc h L:lboratorics, Aug. 1960.
2. Benedict, A. G., Re/iabi/iIJ~confidt>ll("e combinG/ion
for small sample les/.! of aerospace ort/mlnce ilems .
NASA Technical Rcport. pp. 32- 1165. JPL. California
Institute of Tt-chnology. Pasad~na, 1967.
3. Clopper. C . J . and Pearson. E. S., The use ofconfid~nce When /I is small, a $lImpling error may cause eqn (AI) to
or fiducial limits iIlustrak'd in case of the binomial. yield negative estimates. If this occurs. Martz and Waller [4}
Bioml'lrika. 1934. 26,404--413. suggest using the following form of this equation
4. Martz, H. F. and Waller. R. A .• The basics of
Bayesian relia bilit y estimation from attribute test
data. Los Alamos Scientific Laboratory. Report A + B _ (" -
- n
I)("ntR}
tR;- (tRY) _ I
(tR,)l .
(A2)
UC·79p, Febru"ry 1976.
5. Martz, H. F. and Waller. R. A .• Bayesian RI'/iabilil)"
Analysis. John Wiley. 1982. These equations can be applied to processing real life
6. Copas. J . B.. Empirical Bayes methods a nd repeated use data, where" would be the number of test sets for the
of <I standard. Biome/rik". 1972. 59, 349- 360. similar products and R, would be the reliability data from
7. Savchuk, V. ilnd Martz. II ., Bayes reliability ~st imation each sct . h

View publication stats

You might also like