Downloaded from https://www.everyspec.com AMCP 706- IN 2. A statistically sound empirical proof where of the algorithm ?T -failure rate at T 3. A physical model explaining the al- gorithm. T' absolute temperature Unfortunately, most accelerated test tech- XO,7'* = unknown parameters niques do not meet these criteria. They tend to be approximate and require a great kT" , so called activation energy, deal of engineering judgment in the absence where k w Boltzmann constant of precise physical models or statistical techniques. Some dis•.ussions in the litera- To = fixed known temperature ture are vague and ambiguous;, the word "life" may mean the random variable, or Eq.
7-1 can be put in other algebraically may refer to the mean life, or may not mean equivalent forms. anything specific, Beware of using acceler- ation factors from the literature, for specific Where a failure distribution has 2 para- components. The state of the art in corn- meters, it is most common to assume that ponents changes rapidly enough so that one of them is independent of the acceler- failure modes and their acceleration factors ating "stresses". Other assumptions tend can be expected to change in some non- to be intractable, even if more realistic. In appaient way.
the s-normal or lognormal distributions, the median usually is assumed to be a function Temperature is the most common method of the "stresses", and the other parameter of accelerating a test; usually the scale para- to be a constant. meter of the distribution is presumed to No matter what distribution is assumed, follow the Arrhcnius law (see Eq. '-1). it Is essential that the statistiLal uncertaity Voltage can be increased for some kinds of in the results be estimated and clearly capacitors, and power dissipation can be stated -- because this uncertainty is usually increased to shorten the life of many elec- so large as to greatly reduce the impact of tronic components.
Mechanical excitations, the nominal conclusions. In fact, one is such as vibration and shook, are sometimes often tempted to remark, "I could have used. guessed that close without the testsl". For example, using Eq. 7-1 and some reasonably In the parametric approach, the parame- high temperature tests, the uncertainty in tors of the failure distribution are presumed failure rate at operating temperatures might to change in a deterministic fashion with the be a factor of 10 or so. "stress". The functional relationship of the Refs. 3 and 4 give some examples of the "deterministic fashion" Is presumed known, and the purpose of the test is to evaluate application of Eq. 7-1 to real data.
The the parameters in that relationship, maximum likelihood equations for the solution of the problem, and the computer The most common situation is the con- FORTRAN program to effect the solution, stant failure rate and the Arrhenlus tem- are also given there. perature law as shown in Eq. 7-1. Example Nos. 33 and 34 illustrate the (T* T*) procedure. These two examples show how r exp " (7-) grossly misleading it can be to give only To Ti point estimates. 7-10