ML19291A524
| ML19291A524 | |
| Person / Time | |
|---|---|
| Issue date: | 10/18/1977 |
| From: | Abramson L NRC OFFICE OF MANAGEMENT AND PROGRAM ANALYSIS (MPA) |
| To: | Hanauer S NRC OFFICE OF THE EXECUTIVE DIRECTOR FOR OPERATIONS (EDO) |
| Shared Package | |
| ML19291A522 | List: |
| References | |
| FOIA-79-109 NUDOCS 7905160433 | |
| Download: ML19291A524 (3) | |
Text
r
[
s-i a
f
.[f UNITED STATES q
., 7%
NUCLEAR REGULATORY COMMISSION W ASHINGTON,0. C. 20555 nj j,) M i
5
{
K %s p October 18, 1977 I.
1 TO: Stephen H. Hanauer, Technical Advisor Lee R. Abramson, Statistical Adviser h
(,b 'T FROM:
l
SUBJECT:
COMMENTS ON " RELIABILITY DATA ACQUISITION" - 5A (FRAGOLA) i I question the proposed method for converting failure rates for national categories to failure rates for international categories, as it can seriously distort the information contained in the raw I'
data. From page 15, the recom= ended f ailure ra*.e for any inter-national category is the geometric mean of the failure rates for the corresponding national categories. Consequently, if any single national category had a failure race of zero, then the recommended e
failure rate for the international category would also be zero, even if all the other national categories had high failure rates.
But even if this situation does not exist, there are very serious
).'i-problems in using the geometric average. Consider the following hypothetical case with two national categories (component types) comprising the international category:
Failure c3:
106 Hours Rate per I
Component Tvoe Failures of Operation 16 Hours t
A 1
1 1.0 B
10 100 0.1 p-The geometric mean of the failure rates is 0.32, a number which is
- t. ~
[ *-S a gross distortion of the data. Clearly, the Type B data should be
. eighted much more heavily than the Type A data. One way to do this would, of course,, be to use a weightdd geometric mean. However, Q
a better method is simply to use the failure rate for the combined g.
- ~
data, yielding an observed failure rate of 11/101 =.11 for Type A
[-
and B combined. Af ter all, if the international category makes no distinction between Type A and B, then why not calculate the failure fyr '
rate for the international category on this basis?
c.
I recommend calculating the " RECOMMENDED" failure rate for the Et international category by si= ply combining the data for all of the corresponding national categories (this is equivalent to a weighted j.,.
arithmetic mean of the national category failure rates). The failure rates for the various modes can be similarly calculated. Finally,
~
the "HIGH" and " LOW" rates should be taken to be some appropriate (e.g., 95%) confidence bounds based on some reasonable assumptions I
about the underlying failure distribution, r
i e{'
7905160933 er I
h
r 1
.. (
C Other comments follow:
1.
Page 3, 3rd paragraph. The first sentence is a non-sequitur.
The author seems to be confusing syste reliability with
[
component reliability. What he means to say is that the f ailure of some components is a statistically rare event.
(This is an observed fact; it does not follow from the historic reliability of nuclear power plants.)
In the second sentence, " statistically significant samples"
(-
should be replaced by "sufficiently large samples."
2.
Page 4, first paragraph. The author is confusing three concepts:
sample, population, and mixture distribution. First, as a sample from a fixed population is increased, the confidence bounds Secondly, by an " increased" non-homogeneous population, tighten.
the author means a mixture of two different subpopulations (or distributions). The width of a confidence interval for the mean r
of a distribution depends on the variance. The variance of a mixture distribution is always larger than the smallest of the variances of its components, but it may or may not be larger than the largest of the variances of its components, depending on the degree to which the components differ. Hence the g.
confidence bounds for the mean of a mixture distribution will be vider than the narrowest of the confidence bounds for its components y.I and they may or may not be vider than the widest of the component confidence bounds. Of course, calculation of the confidence r
intervals is complicated by the fact that a mixture of Gaussian distributions is not Gaussian unless the components are all b.
identical.
3.
Page 13, Step 3, 2nd paragraph. The third sentence should read L.,-
"The number of components in each of the national categories..."
- *,^
and the heading " POPULATION" in Figure 8 (page 37) should be
- b;;
changed to " NUMBER OF COMPONENTS". In the last sentence, "NPT" 6 hours6.944444e-5 days <br />0.00167 hours <br />9.920635e-6 weeks <br />2.283e-6 months <br /> should be "NFT" and NH should be defined in units of 10 12 as the total number of operating hours of all components in the
,/
[,
nationa.~. category.
N3 "NE" should be "NFC" and NC should be 4.
Page14,firstparagragh. cycles sa the total nu=ber of operating defined in units of 10 cycles of all components in the national category.
g N-w it '
5.
Page 14, 2nd paragraph. The formulas for the failure rates b
should read:
NFT 6 hours6.944444e-5 days <br />0.00167 hours <br />9.920635e-6 weeks <br />2.283e-6 months <br /> AT =
failures per 10
.g
[
AC = 'FC failures per 10 cycles.
V 6
NC I
Lr-i
.t g
M The "N" in the denominators should be omitted; it would be correct only if NH and NC were averages per item.
6.
Page 37, Figure 8.
The number of significant figures in AT and AC should be at most one more than the number of significant i
figures in NFT and NFC, respectively. The percentages for the f
four modes sum to 1.05.
7.
Page 15, Step 4, 2nd paragraph. There is no guarantee that the maximum failure rate will be consistent from country to country.
should be made to define what is meant by " extreme Some attempt I
environmental or other conditions" so that this will have the same meaning to different countries.
8.
Page 17, first paragraph. It might be best to coit Figure 10 on page 39.
It is simply a three element flow diagram repeated r
10 times.
9.
Page 28, Figure 3.
The colu=n of incipient defects lists the means of discovery while the other two columns list the effect
f.
on the system; the table headings should reflect this distinction.
Since an undetected incipient failure cannot enter the data base, the incipient failure rate based on this table is not the same I'?
as the incipient failure rate in the IEEE Guidelines as estimated by Delphi. Furthermore, it should be made clear that incipient and degraded failures which are detected are also corrected; otherwise there vill be double counting of failures if any further deterioration takes place.
s p.
y. $
Eh l*
l5' i
. 6 2
w r, '.,
e 3
A-(
F_:
l
)
f
?'
Y.
e P
F