A Simple Technique to Determine the Exponentiality of Field Failure Data 640575
The formulation of maintenance policy for aircraft and missile systems is usually (unfortunately) set independently of the true distribution of system failures. This is due to several causes:
In-plant life test and product test failure distributions are not always indicative of the field failure distributions, and they never have the same failure rate. Thus, knowledge of both of these factors rarely provides more than vague estimates of the true field failure situation, and therefore is generally neglected.
To gather precise data requires a major effort and usually takes several years (3 or 4) before a proper maintenance policy can be effected.
Even if “good” data were available the samples would have to be quite large and the evaluation of the data would require knowledgable analysis, because the possibility of a decision error is high. Hence, the analysis of such large amounts of data has usually been considered unwieldy and not too necessary.
Since it is now well known that an incorrect maintenance policy can result in a decrease of readiness (particularly true for aircraft and missiles) as well as high costs, it appears that knowledge of a proper maintenance policy would be invaluable.
The object of this paper is to present a technique and a procedure for the analysis of small sample size field failure data to determine if the distribution of failure times is exponential. If the distribution is found to be exponential, then a brief discussion of a technique useful in setting maintenance policy is presented. If the distribution is not exponential then a few useful analysis techniques are presented to locate the non-random times-to-failure.
The main point of the paper is to demonstrate the applicability of the powerful Kolmogorov-Smirnov One Sample Test to field failure data. The simplicity of this test procedure should be sufficient to insure its widespread use in such analysis, but it is also shown how flexible it is by easily locating non-random failures. Presently, such analyses are conducted using the χ2 test. This test is both tedious to apply and relatively inefficient. Furthermore, it is one of the most subjective tests known for application to the analysis of small sample size data.
In this paper, procedures are also presented to assure the proper handling and screening of failure data before analysis is begun.