Department of Defense MIL-HDBK-816 Notice 1 — MIL-HDBK-816 Notice 1 Radiation Hardness Assurance Guidelines Manual PDF

Downloaded from https://www.everyspec.com MIL-HDBK-816 NOTICE 1 4.3.3 Variables tests. Variables LAT tests measure a post-irradiation parameter and apply statistical analysis using approximate or exact probability distributions to describe the variability of that parameter (hence the term variables). Usually pre-irradiation values are also measured and they often figure in the analysis (e.g., when the quantity of interest is the change in the parameter, sometimes referred to also as the parameter “delta”). Sometimes the variable of concern is the stress to failure. 4.3.3.1 Advantages and disadvantages of variables tests.

The major advantage of variables tests is that a high probability requirement may be imposed even if the sample size is only moderate (typically 10-50 parts). A disadvantage is that the measurements and analysis are generally more complex and expensive. The major disadvantage, however, is that assumptions must be made about the probability distribution governing the failure of parts. If the distribution is very well known, then this technique can yield high probabilities with high confidences. However, it is usually difficult to know the “wings” of the distribution accurately.

Therefore, extrapolations to excessively high probabilities (such as 0.999) based on experience with only a few parts must be regarded with skepticism. It is important to note that, for some kinds of electronics effects (e.g., latch-up), there is very little reliable information about the nature of the governing probability distribution. 4.3.3.2 Variables tests depend on tolerance limits. Variables tests depend on tolerance limits which must be compared with the given specifications for the parameter or stress. A lot is rejected if the parameter in question deteriorates beyond an acceptable limit (or if the stress to failure is below the required specification).

The tolerance limits are chosen such that with confidence C, the probability is at least P that parts will not fail the test. 4.3.3.3 Confidence and probability for variables tests. There is a confidence C of rejecting a lot if there is less than probability P that any part in the lot will meet the specified parameter (or stress) tolerance limit. 4.3.3.4 Lognormal distribution. The most commonly assumed distribution for the testing of electronic parts is the lognormal distribution (in the lognormal distribution it is the logarithm of the quantity that is normally distributed). The normal distribution is also frequently used.

Other probability distributions which have been suggested for some circumstances are the Weibull distribution and the extreme value distribution. If enough parts have been tested, the characterization data can be examined, perhaps with a chi-square test, to see which type of distribution fits best. 4.3.3.5 Ramifications to variables tests. As in the case of attribute tests, there are many ramifications to the use of variables tests. In some cases (e.g., step-stress tests) the uncertainty in the measured parameter or stress can be of importance. In other cases a two-sided tolerance limit may be of importance.

Because military device specifications use attribute sample size series tests almost exclusively, this document will not address the ways in which variables tests may be used. It should be recognized, however, that, when systems require very high piecepart survival probabilities, variables tests for lot acceptance may become necessary. 4.3.4 Sampling statistics. The word confidence, as used here, refers strictly to sampling statistics and is the chance of rejecting a lot where there is less than a probability P that any part in the lot can pass the test conditions. The general practice is to take this confidence as the confidence that shipments will be acceptable.

This latter use of the word confidence is a matter that is being discussed in the still somewhat controversial subject of Bayesian statistics and is beyond the scope of this document. The important point is that the latter confidence may only be approximated and involves judgmental decisions. A common practice in hardness assurance is to take the sampling statistics confidence as the confidence that accepted parts will survive. This is an approximation which, though usually valid, may sometimes lead to error.

Perhaps the best justification for such an approximation is that it is usually not the major factor which limits the accuracy and confidence levels associated with system survivability estimates and risk assessments. Often, a more difficult problem, for example, is simulation fidelity (does the test accurately represent the threat environment that is specified for the system). 4.3.5 reasonability of risk assessment results. The major point of the above discussions is to show the approximate nature of risk assessment. Excessively high survival probabilities, such as, for example, a survival probability of 0.999999 (so called 6-nines) may be unreliable for a number of reasons.

First of all, such numbers can only be obtained from extrapolations based on an exact knowledge of the probability distribution. However, because radiation tests use modest numbers of parts and the parts which are tested cannot be used afterwards, knowledge of the assumed probability distribution is never adequate for such extrapolations. Thus, for example, some accidental occurrence during the production of the parts may, with a probability exceeding 1 minus 0.999999, produce a part which does not fit the probability distribution assumed SUPERSEDES PAGE 11 OF MIL-HDBK-816 11