Downloaded from https://www.everyspec.com -MC 706197 the top limit, it wouldn't be too useful. We example, if the desgers whose equipment need an improvement of at least 20 kiprin. 2 " was measured were such that conservative de- Theve is probably little point, then, in running signers put electrolytic capacitors in cool more tests. However, if he says, "All we need paces and careles designers put them in hot is 5 kipslin. improvement," he undoubtedly places, the population of designers does not would want to run more tests to pin down the include those who put very derated electroly- improvement more exactly.
tics in hot places nor those who put mildly s-Confidence is not engineering confi- derated ones in cool places. dence, although the concepts are related. Probably the most controversial situation of samples vs populations concerns the ren- 4-5 GOODNESS-OF-FIT TESTS tionship of cigarette smoking to health. Sam- ples were taken of smokers and nonsmokers, When a particular distribution is assumed etc., but from what population were the peo- to represent a set of data, a natural question ple a statistically random sample?
arises, "How good is the fit of the distribution A more frequently occurring difficulty is to the data?" There are several statistical tests testing a small sample of parts and then ir- that can be performed. Some are peculiar to plicitly hoping that the small sample repre- the distribution itself, and some can be ap- sents the population which will be obtained plied to any distribution. The two most popu- from several suppliers month after month. lar ones for application to any distribution are For really important tests, the engineer the Chi-Square and the Kolmogorov- Smirnov Fo reply important ' ~~tests.
a odcd htaetepsil motn A goodness-of-fit test is equivalent to a effects and then find an appropriate statisti- test of s-significance (par. 4-3) and has all the cian to help with sampling. difficulties associated with s-significance tests. it s pssbleto 4-7 IFR AND DFR DISTRIBUTIONS That difficulty-briefly-is that it is possible to take so few data that it is impossible to reject Sometimes it is difficult to determine a any distribution, and it is possible to take so distribution of lifetimes of a unit. It may, i many data that every distribution will be re- d i neven then, be feasible to decide thet the fail- jected.
ure rate of the unit is always increasing (IFR What is needed is a test for fit that an- - Increasing Failure Rate) or always decreas- swers an engineering question, such as, "If I ing (DFR -* Decreasing Failure Rate). If a dis. use this distribution for interpolation, how tribution is known to be IFR or to be DFR, bad will my answers be?" Unfortunately, such bounds can be put on the failure behavior. tests ace not available. Therefore, a consider- One of these bounds is provided by the Con- able amount of engineering judgment must be stant Failure Rate distribution and its associ- used in reckoning goodness-of-fit. ated relationships.
4-6 SAMPLES AND POPULATIONS For example, the Weibull and Gamma dis- tributions (see Table 3-1 for notation) are In practical situations the population, IFR when the shape parameter 0 is greater about which statistical inferences are to be than 1 and DFR where it is less than 1 Both made, is determined by the method in which have constant failure rates when the shape the sample for testing was drawn. The use of parameter is 1. The s-normal distrbution is historical data is fraught with extreme danger IFR; the lognormal distribution is neither (at this way.
For example, electrolytic capacitors first the failure rate increases, then it de- that were derated to 50 percent or less were more reliable than those derated to, say, 70 A general discussion of IFR and DFR dis- percent of their rating; results like that were tributions is given in Ref. 1; DFR distri- obtained in reliability studies of armed forces butions are discussed in detail in Ref. 1. equipment in the 1950's. Was this sample Bounds on reliahility parameters are given in taken from all kinds of designers, or was it Refs. 2-5. Refs. 6, 7 discuss the conditions taken from only a subset of designers? For under which systems: 4-3