Beruflich Dokumente
Kultur Dokumente
Herbert D. Peckham*
HOLEX incorporated
Hollister, California
628
SUPPLEMENT TO IEEE TRANSACTIONS ON AEROSPACE / JUNE 1965
The type and quantity of explosives in a given EED will give a factor K which when multiplied by s
are varied to meet specific design requirements. and added to x will produce the required stimulus
level. ] n this particular instance the value of K
The characteristics of EED's make them particu- is 3. 3( . The desired stimulus level is therefore
larly well suited to aerospace applications. Their 2. 0 + (3. 3) (0. 1) = 2. 33 amperes . The final
energy potential to weight ratio is extremely high. result of the sensitivity test is the statement that
One of their outstanding characteristics is ex- assuming normal distribution of data, there is a
treme simplicity with resultant high reliability. 9 5 % confidence level that at lea st 9 9. 9 % of the
Properly designed, and utilizing carefully EED's will function at a stimulus level of 2. 33
selected explosives, EED's can be subjected to a amperes.
variety of severe environmental stresses and still
function reliably as designed. PART II: METHODS OF SENSITIVITY TESTING
629
SUPPLEMENT TO IEEE TRANSACTIONS ON AEROSPACE / JUNE 1965
levels can be predetermined and the requisite exists even assuming that the variate being inves-
number of parts with each load can be assembled. tigated has a normal distribution - a requirement of
Then, once testing is commenced, no changes in the Bruceton method. If there is reason to question
load levels are required. On the other hand, if the normalcy of the input data, the results of the
the load level of each shot was determined by the Bruceton method involving extrapolation are doubly
results of the previous shot, as is the case in suspect.
some methods of sensitivity testing, the testing
would be unreasonably drawn out. The Probit RUN DOWN METHOD
method will generally require more parts to pro-
duce the same results than other more efficient The Run Down method was developed at the Naval
methods of sensitivity testing. A basic consider- Ordnance Laboratory, White Oak, Maryland, and is
ation is that the method is based upon normalcy described in a Bureau of Naval Weapons military
of the variate being tested. specification. (5) The novel features of the Run
Down method are that the method assumes a very
BRUCETON METHOD conservative EED probability distribution -- the
log logistic distribution, and generates test data
The Bruceton (or up and down) method of sensiti- considerably further out on the tails of the distri-
vity testing was developed specifically for ord- bution than is the case in Bruceton testing.
nance testing. The majority of all EED sensiti-
vity testing carried out at the present time An initial Bruceton test is conducted using 20 parts.
utilizes this method. Crow(3) and Dixon(4) The purpose is to gain a rough idea of the charac-
have given excellent descriptions of the method teristics of the distribution. Based upon the out-
complete with calculation procedures. put of the initial Bruceton test, two firing levels
are computed -- level 1 (x + 0. 4s), and level 2
The Bruceton test is carried out using equally (x + 1, 3s). Fifty parts are tested at level 1. If
spaced test levels that are chosen before the 5 or fewer fires are observed testing is continued
test commences. The test is sequential in that until 130 parts have been tested. Then a new level
the test level for a given shot is determined by is computed (5x + 0. 2s) at which 50 parts are tested.
the outcome of the previous shot. If a given shot In the usual case in which more than 5 fires are
fires, the next shot is made at the next lower test observed at level 1, 130 parts are tested at level
level. If a given shot does not fire, the subse- 2. The total number of parts required is 200.
quent shot is made at the next higher test level.
Table 2 shows the results of a Bruceton test car- The data obtained above is analyzed for a reliabi-
ried out on 36 electro-explosive cartridges. The lity of 99. 0% by the Berkson technique. (6) A
test stimulus is firing current and the test levels method of calculating confidence limits has been
are spaced 0.05 amperes apart. Based upon the described by Working and Hotelling. (7) The out-
results in Table 2, the required calculations yield put of the Run Down test is, as before, a stimulus
x = 1.41 amperes, and s = 0.036 amperes. level which corresponds to a specified response
and confidence level.
Test
Level Response to Test An important fact about the Run Down method is
(Amperes) (x = Fire, o = No-Fire) that it generates data further out on the probability
distribution than is the case in Bruceton testing.
1.50 x It is, therefore, less subject to extrapolation
1.45 x x x X o x x x x x x x x x errors than the Bruceton method. This method,
1.40 x o o xoo oooooooo x x although an improvement over the Bruceton method,
1.35 0 0 ox does not solve the situation completely. The Run
1.30 o Down method can be used out to a reliability of
99. 0% as described in the Military Specification.
Table 2. Typical Bruceton Test Results However, typical EED reliability requirements
today are 99.99% response at a confidence level of
95%. Therefore, even though testing is carried out
The nature of the Bruceton method causes testing further out on the tails than is the case using
to be concentrated about the mean. It is therefore Bruceton testing, long extrapolations are still re-
a very efficient estimator of the mean -- more so quired to reach current reliability requirements.
than any other method. It is not a good estimator
of the standard deviation. Since the testing is PART III: PROBLEM AREAS
concentrated about the mean, little or no data is
gathered in the tails of the distribution. The PROPER PROBABILITY DISTRIBUTION FUNCTIONS
result is a tendency to underestimate the standard
deviation. This is particularly dangerous in the A central and critical element in all the methods of
case of EED's because it makes them appear to be sensitivity testing discussed thus far is the appro-
more reliable than they really are at the rated priate probability function which describes the EED
firing current, and safer than they really are to the being tested. Both the Bruceton and Probit methods
hazard of low stimulus initiation. It is important require that the variate being tested be normally
to note that this weakness of the Bruceton method
630
SUPPLEMENT TO IEEE TRANSACTIONS ON AEROSPACE / JUNE 1965
distributed. The Run Down method assumes a log closely connected with the problem of EED relia-
logistic distribution. bility esitmation. Obviously, as the reliability
requirements get higher, the number of parts allo-
Considering the number of EED's currently em- cated to prove the reliability should increase pro-
ployed in aerospace applications, a surprisingly portionately. However, experience has proven
small amount of work has been done to determine that this is not always the case. Where the indi-
the appropriate statistical model for EED's. At vidual unit cost is high, there have been cases
the Franklin Institute Laboratories, 4, 362 carbon where the sample size for reliability testing has
bridge EED's were tested to determine the statis- actually been decreased at the higher reliability
tical model, (8) and more recently 7, 890 hot-wire levels.
EED's were expended in a statistical model test
at the Naval Ordnance Laboratory. (9) The results
of these tests indicate that the proper variate is It is understandable that the budget conscious
the logarithm of the firing current, and that this engineer is loath to allocate large numbers of ex-
variate is distributed according to the logistic pensive parts to be destructively tested in sensi-
distribution. This information is of little use to tivity testing. However, if the mission of the EED
the contemporary reliability engineer because the is sufficiently important to warrant a high relia-
state of the art has advanced far past the carbon bility and confidence level, an expenditure of the
bridge EED or the type of hot-wire EED which was necessary number of parts to demonstrate the re-
tested at the Naval Ordnance Laboratory. liability is certainly justified. Failure to properly
verify the reliability requirements is a sure path to
The problem is that there is no assurance that the mission failure.
more sophisticated, modem EED's follow a log-
logistic distribution, or even if they have a The requirement for reliability has been steadily
generic "best" distribution. It is quite possible increasing. In the late 1950's, many of the aero-
that the statistical model appropriate to one type space projects did not require reliability and con-
of modem EED would not be suitable to another. fidence figures to be associated with a given
The point is that this is a critical area of uncer- stimulus to an EED. The picture has changed how-
tainty which needs to be resolved if valid EED ever. In 1961-1962, reliability figures of 99.9%
reliability estimates are to be made. were common. Now, the usual reliability require-
ment on aerospace EED's is 99.99% at a confidence
NORMALIZING TECHNIQUES level of 95%. There have been cases in which the
reliability figure has been set at 99. 999% and
In cases where a sensitivity test is used which is higher. To put this in the proper focus, it must be
based upon normalcy of data, but the variate being remembered that using the most common method of
tested is not normally distributed, it is usually sensitivity testing, the Bruceton method, that
possible to find a transform which will generate a reliability estimates past the 99% point must be
new variate which is, or at least is approximately, considered suspect because of the requirement for
normally distributed. For example, it is held by normalcy of data. This, coupled with low sample
some authors that EED critical firing current is not size, begins to lend a suspicious cast upon much
normally distributed, but the new variate gen- of the contemporary EED reliability data. One
erated by taking the logarithm of the critical firing suspects that reliability requirements may be satis-
current is normally distributed. The log transform fied only on paper. Also, looking at the sky-
is relatively common and indeed is the subject of rocketing reliability requirements, one wonders if
an entire book. (10) they reflect true system reliability requirements or
are the result of several stages of reliability
There are many other transformations which will "padding" accomplished to conceal uncertainty on
make almost any distribution normal in the vicinity the part of the design engineer.
of the mean, but to guarantee normalcy of data in
the tails of the distribution is quite another PART IV: REMEDIAL ACTION
matter. In cases in which normalcy dependent
sensitivity testing is used to estimate extreme PROBABILITY DISTRIB UTIONS
reliability points, nothing short of an extensive
statistical model investigation involving perhaps A thorough knowledge of the appropriate probability
thousands of observations will suffice.jl) It distribution function (or functions) is essential
must be emphasized therefore that too much con- before meaningful discussion can take place about
fidence cannot be placed on a normalizing tech- EED reliability. As indicated previously, there is
nique to permit the use of the Bruceton or Probit a paucity of information about EED probability dis-
method unless it can be proven that the tails of tribution functions and what is available concerns
the distribution have been normalized. This is obsolete devices. Most certainly, the reason why
particularly important when estimating extreme more statistical model investigations have not been
reliability points as is the case most often today. made is because of the cost involved. Very rarely
can the statistical investigator procure the several
SAMPLE SIZE AND RELIABILITY REQUIREMENTS thousand devices needed. The cost of 5, 000
sophisticated EED's similar to those in use on con-
Sample size and reliability requirements are temporary and planned aerospace applications
631
SUPPLEMENT TO IEEE TRANSACTIONS ON AEROSPACE / JUNE 1965
would cost at least $50, 000. 00. It is a rare convergence rate upon the desired response point.
budget that can include figures of this magnitude This knowledge is not presently available. Also,
for reliability investigations. workable cut off rules need to be developed to stop
the testing process. Methods of introducing confi-
An easier approach is to characterize modem EED's dence levels into the calculations are available,
by construction and material. Once generic but need definition towards the specific problem of
'types" have been defined which permit the ready EED sensitivity testing.
identification of a given EED as to its type, the
task is to determine the appropriate statistical It should be noted that even though the method is
model of each type. With some ingenuity, an EED non-parametric, a knowledge of the probability
could be designed to serve as a statistical test distribution involved can make the method more
vehicle at a fraction of the cost of a flight EED. efficient and require less parts. By generating
The reduction in cost would take place through computer programs which can simulate the method
the elimination of any features superfluous to the and the probability distributions involved, large
requirement of -sensitivity testing. The test amounts of experience can be acquired relatively
vehicle should be designed to permit simulation of quickly. With the proper experience, the relia-
all the "type" EED's. bility engineer can estimate in advance how many
parts will be required to locate a given response
The first step that should be taken, then, is to point. This would inject a refreshing note into the
undertake a serious investigation of the appro- combined process of specifying and proving re-
priate statistical models which describe the most liability.
modern EED's. This must be done before any pre-
cision can be inserted into the process of EED The second step which needs to be taken, then, is
reliability estimation. With a properly designed to develop efficient methods of non-parametric
EED to serve as a test vehicle, this could be sensitivity testing and couple the new tests with a
accomplished with a reasonable budget. knowledge of the appropriate statistical models of
EED's.
NON-PARAMETRIC TESTING METHODS
CONCLUSION
All of the sensitivity testing methods discussed in
this paper thus far are parametric methods. That Mission success depends critically upon all the
is, they depend critically upon the assumption of components of a system functioning reliably as
various parameters. The outcome of the test is designed. For electro-explosive devices (EED' s)
directly related to the degree of accuracy with the verification of this reliability must be done
which the original assumptions are verified. Thus, with some form of sensitivity testing. Insufficient
there is a built in restriction on the tests which attention to valid demonstrations of EED reliability,
can cause serious errors if ignored. or failure to observe important restrictions on
sensitivity tests, can be an invitation to system
The non-parametric tests are completely different. failure or can lead to rejection of reliable parts.
They do not depend upon the assumption of any
particular distribution or parameter of a distribu- Of the methods of sensitivity testing currently
tion. There are mild requirements which must be available, the one used most often is the Bruceton
met but these are very easy to satisfy. Most of method. The Bruceton method requires that the
the existing non-parametric testing methods variate being investigated be normally distributed.
suitable for sensitivity testing are variations of There is relatively little information available con-
the Robbins-Monroe method. (2) cerning the proper statistical model for EED's.
Thus, the employment of a sensitivity test based
The basic equation of the method which produces upon normalcy of data against unknown probability
successive test levels is: distributions must be considered suspect. The use
of normalizing transforms to normalize the variate
Xn =
Xn-l - 9
Yn-l - zj (1) are effective in the region of the mean, but are not
valid for extreme response points.
In equation (1) xn is the test stimulus for the nth
test, xn-l is the test stimulus for the (n - 1)St
The solution to the dilemma is twofold. First,
test, C is an undetermined constant, f(n) is an serious investigations must be made to determine
undetermined function of n, Yn-l is the response the appropriate statistical model (or models) to
of the (n - l)st test (for success y = 1, for accurately describe modern and sophisticated EEIYs.
failure y = 0), and z is the decimal which repre- Second, non-parametric tests need to be developed
sents the response point desired. whi ch will function efficiently even in the face of
deviations from expected statistical distributions.
The Robbins-Monroe method has been in mathe-
matical literature for some time but relatively It is quite possible, and indeed probable, that
little use has been made of it from the engineering EED's scheduled for use have passed all reliability
standpoint. The reason is that one must make the requirements which have been specified but which
optimum choice of C and f(n) to produce maximum have true reliability figures differing widely from
632
SUPPLEMENT TO IEEE TRANSACTIONS ON AEROSPACE / JUNE 1965
REFERENCES
(1) Lloyd, D.K. & Lipow, M., "Reliability:
Management, Methods, and Mathematics,"
Prentice Hall Inc., Englewood Cliffs, N.J.,
pp. 496-497, 1962.
(2) United States Department of Commerce Hand-
book 91, "Experimental Statistics, " U. S.
Government Printing Office, Section 10-3,
19 63.
(3) Crow, E.L., Davis, F.A., & Maxfield, M.W.,
"Statistics Manual, " Dover Publications,
Inc., NewYork, Section 4.5.1, 1960.
(4) Dixon, W.J. & Massey, F.J., "An Introduc-
tion to Statistical Analysis, " McGraw Hill,
New York, Chapter 19, 1955.
(5) Department of the Navy, Bureau of Naval
Weapons, MIL-I-23659 (Wep), "Military
Specification, Design & Evaluation of
Electric Initiators, " 18 Marhc 1963.
(6) Berkson, J., "A Statistically Precise and
Relatively Simple Method of Estimating the
Bio-assay and Quantal Response, Based on
the Logistic Distribution Function, " Journal
of The American Statistical Association, 58,
pp. 565 - 569, September 1953.
(7) Working & Hotelling, Proceedings of the
American Statistical Association, p 73 of
supplement to Vol. 24, Series No. 165A.
(8) Hammer, C., "Statistical Methods in
Initiator Evaluation," Franklin Institute
Laboratories for Research and Development,
Interim Report No. I-1804-1, May 1, 1955.
(9) Hampton, L.D., &Ayres, J.N., "Charac-
terization of Squib Mk 1 Mod 0, Determina-
tion of the Statistical Model," NAVWEPS
Report 7347, January 30, 1961.
(10) Aitchison, J. & Brown, J.A.C., "The Log-
normal Distribution, " Cambridge University
Press, Cambridge, England, 1957.
(11) Dixon, W.J. & Massey, F. J., op cit, par
19-2.
(12) Robbins, H., & Monro, S., "A Stochastic
Approximation Method, " Annals of Mathe-
matical Statistics 22, '51, pp. 400-407.
633