With product reliability demonstration test planning and execution weighing heavily on cost, availability and schedule factors, Bayesian methods offer an intelligent way of incorporating engineering knowledge based on historical information into data analysis and interpretation, resulting in an overall more precise and less resource intensive failure rate estimation. This talk consists of three parts
1. Introduction to Bayesian vs Frequentist statistical approaches
2. Bayesian formalism for reliability estimation
3. Product/component case studies and examples
2. ASQ Reliability Division
ASQ Reliability Division
English Webinar Series
English Webinar Series
One of the monthly webinars
One of the monthly webinars
on topics of interest to
reliability engineers.
To view recorded webinar (available to ASQ Reliability
( y
Division members only) visit asq.org/reliability
To sign up for the free and available to anyone live webinars
To sign up for the free and available to anyone live webinars
visit reliabilitycalendar.org and select English Webinars to
find links to register for upcoming events
http://reliabilitycalendar.org/The_Rel
iability Calendar/Webinars_‐
y_ /
_English/Webinars_‐_English.html
3. Bayesian Methods in Reliability Engineering
ASQ Reliability Division Webinar Program
Nov 15th 2012
Charles H. Recchia, MBA, PhD
Quality Support Group, Inc
http://www.qualitysupportgroup.com/
http://www qualitysupportgroup com/
4. BAYESIAN METHODS IN RELIABILITY ENGINEERING
With product reliability demonstration test planning and execution interacting
heavily with cost, availability and schedule considerations, Bayesian methods
heavily with cost availability and schedule considerations Bayesian methods
offer an intelligent way of incorporating engineering knowledge based on
y p g
historical information into data analysis and interpretation, resulting in an
overall more precise and less resource intensive failure rate estimation. This talk
consists of three parts
Introduction to Bayesian vs Frequentist statistical approaches
Bayesian formalism for reliability estimation
Product/component case studies and examples
Charles Recchia has more than two dozen years of fundamental research, technology/product
development, and management experience with a special focus on reliability statistics of complex
systems. He earned a doctorate in Condensed Matter Physics from Ohio State University, and a Master
of Business Administration degree from Babson College. Dr. Recchia accrued reliability engineering
fB i Ad i i i d f B b C ll D R hi d li bili i i
expertise at Intel, MKS Instruments and Saint‐Gobain Innovative Materials R&D, has served as adjunct
professor at Wittenberg University, and is author of numerous peer‐reviewed technical papers and
patents. He is a senior member of ASQ, the American Physical Society, and serves on the Advisory
f f
Committee for the Boston Chapter of the IEEE Reliability Society.
11/15/2012 ASQ RD Webinar 2
5. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 3
7. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 5
8. A d
Agenda
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 6
9. When reliability follows the exponential TTF
When reliability follows the exponential TTF model (eg the flat
the flat
constant failure rate portion of Bathtub Curve):
Classical Framework
– The MTBF is one fixed unknown value ‐ there is no “probability” associated
with it
ith it
– Failure data from a test or observation period allows you to make
inferences about the value of the true unknown MTBF
– No other data are used and no “judgment” ‐ the procedure is objective and
based solely on the test data and the assumed HPP model
Bayesian Framework
Bayesian Framework
– The MTBF is a random quantity with a probability distribution
– The particular piece of equipment or system you are testing “chooses” an
MTBF from this distribution and you observe failure data that follow an
MTBF f hi di ib i d b f il d h f ll
HPP model with that MTBF
– Prior to running the test, you already have some idea of what the MTBF
probability distribution looks like based on prior test data or an consensus
engineering judgment
11/15/2012 ASQ RD Webinar 7
10. exponential distribution
ti l di t ib ti
“non‐intuitive”
non intuitive
Brains wired with
Planet = Earth
Normal Distribution σ ~ 0.1μ 6 draw sample.
Population mean height 5’11”
Sample mean = ______
8
5’9” 6’3” 5’10” 5’9” 6’0” 5’9”
11. ti l di t ib ti
exponential distribution
Planet = Laitnenopxe
6 draw sample.
Population mean height 5’11”
Sample mean = ______
11/15/2012 ASQ RD Webinar 9
4’11” 13’8” 7’11” 5’9” 2’5” 4”
13. For and Against use of Bayesian Methodology
g y gy
PROs CONs
Uses prior information ‐ thi
U i i f ti this Prior information may not be
Pi i f ti tb
"makes sense“ accurate ‐ generating misleading
Less new testing may be
Less new testing may be conclusions
needed to confirm a desired Way of inputting prior information
MTBF at a given confidence
MTBF at a given confidence (choice of prior) may not be correct
(choice of prior) may not be correct
Confidence intervals are really Customers may not accept validity of
intervals for the (random) MTBF
intervals for the (random) MTBF prior data or engineering judgements
prior data or engineering judgements
‐ sometimes called "credibility Risk of perception that results
intervals“ aren't objective and don't stand by
j y
themselves
11/15/2012 ASQ RD Webinar 11
14. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 12
15. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 13
16. Toy Hyperbolic Example
y yp p
First day on the job as reliability engineer, you overhear three colleagues
debating the MTBF for a product.
Evidently the engineer you are replacing had kept all his data on his now‐
destroyed C:drive and all that remains is “word of mouth” among his three
remaining coworkers.
Waloddi: “I remember seeing 500 seconds written down
on his whiteboard. I can still see it in my head.”
Gertrude: “No W, that was 100 seconds! His
handwriting was atrocious, but that definitely was a 1
not a 5.
not a 5 ”
Taiichi: “Agree with Gertrude. It was 100 seconds.”
2 against 1. That is all you have to go by. Your manager needs an answer by end
of day. The ink on your badge hasn’t even dried yet.
fd Th i k b d h ’t di d t
What shall you do? They measure product MTBF in seconds?
11/15/2012 ASQ RD Webinar 14
17. Bayesian Core Idea
Bayesian Core Idea
What you knew
What you knew
before WYKB. New Data
“Prior”
Best possible update of WYKB
adjusted by the New Data.
adjusted by the New Data
“Posterior”
11/15/2012 ASQ RD Webinar 15
18. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 16
19. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
There is just enough time 1.43E‐18
There is justLikenough) time21 1 43E 18
Likelihood Π f (
lih d (t 5.49E‐21
5 49E i
before end of day to collect 6
average {t }
average {t } = 317 i y
i TTF data t (sec) f (t ) f (t )
time‐to‐fail (TTF) data points.1.53E‐03
1 133
i
2.65E‐03
1 i 2 i
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
Let’s do that. 5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 17
20. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 18
21. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
How likely is this data?
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i ) 5 49E 21 what λ 18
Likelihood Depends on
Depends on what43E is!
(t 5.49E‐21 1.43E‐18
1
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 19
22. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 20
23. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 21
24. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 22
25. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
Shouldn t these (sec)
Shouldn’t these
MTTF 450 100 500
sum to 1 if they
are exhaustive
Prior g (λ ) 0.667 0.333
possibilities? P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 23
26. λ KOtG λ 1 λ 2
Failure Rate λ (1/sec) 0.0022 0.0100 0.0020
MTTF (sec) 450 100 500
Prior g (λ ) 0.667 0.333
P t i g (λ | t i )
Posterior 0.204
0 204 0.796
0 796
Prior*Likelihood 3.66E‐21 1.43E‐20
Lik lih d Π f ( i )
Likelihood (t 5.49E‐21
5 49E 21 1.43E‐18
1 43E 18
average {t }
average {t i } = 317
i TTF data t i (sec) f 1(t i ) f 2(t i )
1 133 2.65E‐03 1.53E‐03
2 888 1.39E‐06 3.39E‐04
3 619 2.05E‐05 5.80E‐04
4 8 9.23E‐03
9 23E 03 1.97E‐03
1 97E 03
5 97 3.78E‐03 1.65E‐03
6 157 2.08E‐03 1.46E‐03
11/15/2012 ASQ RD Webinar 24
28. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 26
29. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
Conjugate Priors
• Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 27
31. Mean λave = a/b
Variance σ2 = a/b2
In hierarchical Bayesian models, these hyperparameters
will be represented as distributions with priors/posteriors, etc
will be represented as distributions with priors/posteriors etc
and have hyperparameters of their own
11/15/2012 ASQ RD Webinar 29
32. Bayesian assumptions for the gamma
y p f g
exponential system model
1. Failure times for the system under investigation can be adequately
modeled by the exponential distribution with constant failure rate.
2. The MTBF for the system can be regarded as chosen from a prior
distribution model that is an analytic representation of our
distribution model that is an analytic representation of our
previous information or judgments about the system's reliability.
The form of this prior model is the gamma distribution (the
conjugate prior for the exponential model).
The prior model is actually defined for λ = 1/MTBF.
3. Our prior knowledge is used to choose the gamma parameters
3 O i k l d i dt h th t
a and b for the prior distribution model for λ. There are a number
of ways to convert prior knowledge to gamma parameters.
y p g g p
11/15/2012 ASQ RD Webinar 30
33. Gamma prior parameter method 1
G i t th d 1
1. If you have actual data from previous testing done on the
system (or a system believed to have the same reliability as the
system (or a system believed to have the same reliability as the
one under investigation), this is the most credible prior
knowledge, and the easiest to use. Simply set the gamma
knowledge, and the easiest to use. Simply set the gamma
parameter a equal to the total number of failures from all the
p e ous da a, a d se e pa a e e equa o e o a o a
previous data, and set the parameter b equal to the total of all
the previous test hours.
11/15/2012 ASQ RD Webinar 31
34. Gamma prior parameter method 2
Gamma prior parameter method 2
2. A consensus method for determining
2 A consensus method for determining a and b that works well is the following:
that works well is the following:
Assemble a group of engineers who know the system and its sub‐components well
from a reliability viewpoint.
A. Have the group reach agreement on a reasonable MTBF they expect the system to have.
They could each pick a number they would be willing to bet even money that the system
would either meet or miss, and the average or median of these numbers would be their
50% best guess for the MTBF. Or they could just discuss even‐money MTBF candidates until
a consensus is reached.
p p g , g g y p
B. Repeat the process again, this time reaching agreement on a low MTBF they expect the
system to exceed. A "5%" value that they are "95% confident" the system will exceed (i.e.,
they would give 19 to 1 odds) is a good choice. Or a "10%" value might be chosen (i.e., they
g ) p
would give 9 to 1 odds the actual MTBF exceeds the low MTBF). Use whichever percentile
choice the group prefers.
C. Call the reasonable MTBF MTBF50 and the low MTBF you are 95% confident the system will
exceedMTBF05. These two numbers uniquely determine gamma parameters a and b that
. These two numbers uniquely determine gamma parameters that
have percentile values at the right locations
Called the 50/95 method (or the 50/90 method if one uses MTBF10 , etc.)
11/15/2012 ASQ RD Webinar 32
35. Gamma prior parameter method 3
G i t th d 3
3. Weak Prior Obtain consensus is on a reasonable
expected MTBF, called MTBF Next however the
expected MTBF called MTBF50. Next, however, the
group decides they want a weak prior that will change
rapidly, based on new test data. If the prior parameter
rapidly based on new test data If the prior parameter
"a" is set to 1, the gamma has a standard deviation
equal to its mean, which makes it spread out, or
equal to its mean which makes it spread out or
"weak". To set the 50th percentile we have to
choose b = ln 2 × MTBF50
= ln 2 ×
Note: During planning of Bayesian tests, this weak prior is
actually a very friendly prior in terms of saving test time.
11/15/2012 ASQ RD Webinar 33
36. Comments
C t
Many variations are possible, based on the above
three methods. For example, you might have prior
three methods For example you might have prior
data from sources that you don't completely trust. Or
you might question whether the data really apply to
i ht ti h th th d t ll l t
the system under investigation. You might decide to
"weight" the prior data by .5, to "weaken" it. This can
be implemented by setting a = .5 x the number of
p y g
fails in the prior data and b = .5 times the number of
test hours. That spreads out the prior distribution
test hours. That spreads out the prior distribution
more, and lets it react quicker to new test data.
11/15/2012 ASQ RD Webinar 34
37. New data is collected …
N d t i ll t d
New information is combined with the gamma prior model to
p
produce a gamma posterior distribution.
g p
After a new test is run with
T additional system operating hours, and
r new failures,
The resultant posterior distribution for failure rate λ remains
gamma (conjugate remember?), with new parameters
a' = a + r
'
b' = b + T
b' b + T
11/15/2012 ASQ RD Webinar 35
39. Example
p
• A group of engineers, discussing the reliability of a new
piece of equipment decide to use the 50/95 method to
equipment,
convert their knowledge into a Bayesian gamma prior.
Consensus is reached on a likely MTBF50 value of 600 hours
and a low MTBF05 value of 250.
• RT is 600/250 = 2.4. Using software to find the root of a
/ g
univariate function, the gamma prior parameters were
found to be a = 2.863 and b = 1522.46. The parameters will
have (
h (approximately) a probability of 50% of b i b l
i t l ) b bilit f f being below
1/600 = 0.001667 hours‐1 and a probability of 95% of being
below 1/250 = 0 004 hours‐1. (The probabilities are based on
0.004
the 0.001667 and 0.004 quantiles of a gamma distribution
with shape parameter a = 2.863 and scale parameter b =
p p p
1522.46 hours)
11/15/2012 ASQ RD Webinar 37
40. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
Conjugate Priors
• Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti
Question and Answer
dA
11/15/2012 ASQ RD Webinar 38
41. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Question and Answer
Q ti dA
11/15/2012 ASQ RD Webinar 39
42. Bayesian test plan
y p
Gamma prior parameters a and b have already been determined. Assume we have a given MTBF
objective, M, and a desired confidence level of 100×(1‐ α). We want to confirm the system will have
an MTBF of at least at the 100×(1‐ α) confidence level. Pick a number of failures, r, that we can
an MTBF of at least M at the 100×(1 α) confidence level Pick a number of failures r that we can
allow on the test.
We need a test time T such that we can observe up to r failures and still "pass" the test. If the test
time is too long (or too short), we can iterate with a different choice of the test ends, the posterior
gamma distribution will have (worst case ‐ assuming exactly r failures) new parameters of a ' = a + r,
b' = b + T and passing the test means that the failure rate λ1‐ α, the upper 100×(1‐ α) percentile for the
posterior gamma, has to equal the target failure rate 1/M.
By definition, this is G ‐1(1‐ α; a', b'), with G ‐1 denoting the inverse of the gamma CDF distribution .
We can find the value of that satisfies 1(1‐ α;a b') = 1/M by trial and error. However, based on
We can find the value of T that satisfies G ‐1(1 α;a', b ) = 1/M by trial and error However based on
the properties of the gamma distribution, it turns out that we can calculate T directly by using
T = M×(G ‐1(1‐ α; a', 1)) ‐ b
11/15/2012 ASQ RD Webinar 40
43. Special Case: a = 1 (The "Weak" Prior)
p ( )
When the prior is a weak prior with a = 1, the Bayesian test is always
shorter than the classical test.
There is a very simple way to calculate the required Bayesian test time
when the prior is a weak prior with a = 1. First calculate the
classical/frequentist test time Call this Tc. The Bayesian test time T is
test time. Call this The Bayesian test time is
just Tc minus the prior parameter b (i.e.,T = Tc ‐ b). If the b parameter
was set equal to (ln 2) × MTBF50(where MTBF50 is the consensus
choice for an "even money" MTBF), then T = T
choice for an "even money" MTBF) then T Tc ‐ (ln 2) × MTBF50This
2) × This
shows that when a weak prior is used, the Bayesian test time is always
less than the corresponding classical test time. That is why this prior is
p g y p
also known as a friendly prior.
This prior essentially sets the order of magnitude for the MTTF
11/15/2012 ASQ RD Webinar 41
44. Calculating a Bayesian Test Time
g y
A new piece of equipment has to meet a MTBF requirement of 500 hours at 80
% confidence. A group of engineers decide to use their collective experience to
determine a Bayesian gamma prior using the 50/95 method described
determine a Bayesian gamma prior using the 50/95 method described
in Section 2. They determine that the gamma prior parameters are a = 2.863
and b = 1522.46 hrs.
Now they want to determine an appropriate test time so that they can confirm
N th tt d t i i t t t ti th t th fi
a MTBF of 500 with at least 80 % confidence, provided they have no more than
two failures (r = 2).
We obtain a test time of 1756.117 hours using 500×(G ‐1(1‐0.2; 2.863+2, 1)) ‐
1522.46
To compare this result to the classical test time required, which is 2140 hours
To compare this result to the classical test time required, which is 2140 hours
for a non‐Bayesian test. The Bayesian test saves about 384 hours, or an 18 %
savings. If the test is run for 1756 hours, with no more than two failures, then
an MTBF of at least 500 hours has been confirmed at 80 % confidence.
an MTBF of at least 500 hours has been confirmed at 80 % confidence.
If, instead, the engineers had decided to use a weak prior with an MTBF50 of
600, the required test time would have been 2140 ‐ 600 × ln 2 = 1724 hours
11/15/2012 ASQ RD Webinar 42
45. Post‐Test Analysis Example
P t T tA l i E l
• A system has completed a reliability test aimed at
confirming a 600 hour MTBF at an 80% confidence
confirming a 600 hour MTBF at an 80% confidence
level. Before the test, a gamma prior with a = 2, b =
1400 was agreed upon, based on testing at the
1400 was agreed upon, based on testing at the
vendor's location. Bayesian test planning
calculations, allowing up to 2 new failures, called
calculations, allowing up to 2 new failures, called
for a test of 1909 hours.
• When that test was run there actually were exactly
When that test was run, there actually were exactly
two failures. What can be said about the reliability?
The posterior gamma CDF has parameters a = 4
The posterior gamma CDF has parameters a' = 4
and b' = 3309.
11/15/2012 ASQ RD Webinar 43
46. Bayesian solutions for arbitrary F(t)
B i l ti f bit F(t)
What about Weibull
Wh t b t W ib ll or other non‐exponential variable failure rate TTF
th ti l i bl f il t TTF
distributions?
3.50E‐02
3.00E‐02
2.50E‐02 β = 0.6
2.00E‐02
2 00E 02 β=08
0.8
g(λ, β|data) β = 1.0
1.50E‐02
β = 1.2
1.00E‐02 β = 1.4
5.00E‐03 β = 1.6
0.00E+00
0.0000 0.0020 0.0040 0.0060 0.0080
λ (1/sec)
Conjugate priors only exist for Weibull when a subset of hyperparameters are
Conjugate priors only exist for Weibull when a subset of hyperparameters are
known. MCMC and Gibbs methods exist for sampling from higher dimensional
posteriors CDFs in multiple dimensions not as straightforward.
11/15/2012 ASQ RD Webinar 44
47. References and Further Reading
R f d F th R di
• NIST/SEMATECH e‐Handbook of Statistical Methods,
http://www.itl.nist.gov/div898/handbook/, April (2012)
• Statistical Methods for Reliability Data, WQ Meeker and LA
Escobar (1998)
• Applied Reliability, 2nd edition, PA Tobias and DC Trindade
(1995)
• Bayesian Reliability Analysis, HF Martz and RA Waller (1982)
• Methods for Statistical Analysis of Reliability and Life Data,
NR Mann, RE Schafer, and ND Singpurwalla (1974)
NR M RE S h f d ND Si ll
• Bayes is for the birds, RA Evans, IEEE Transactions on
Reliability R‐38, 401 (1989).
R li bilit R 38 401 (1989)
11/15/2012 ASQ RD Webinar 45
49. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Question and Answer
Q ti dA
11/15/2012 ASQ RD Webinar 47
50. Agenda
A d
• Bayesian vs. Frequentist Comparison
• Preliminary Example
• Conjugate Priors
Conjugate Priors
• Test Time Examples
Test Time Examples
• Q ti dA
Question and Answer
11/15/2012 ASQ RD Webinar 48