an estimate of the standard deviation of the historical estimate (of this difference). The critical values are chosen, using historical data, to balance errors of identifying a process in need of correction when it is functioning fine against the cost of letting a process pass that is in need of correction. (For details, see Kalbfleisch et al., 1991 and Wu and Meeker, 2002.)
A second important application of field performance data is the prediction of future warranty or total maintenance costs (the second possibility currently being more relevant to DoD systems). Clearly, information on the rate of field failures of various types could be extremely useful for estimating field maintenance, repair, and component replacement costs.
A third use of field performance data is to establish a “transfer function” between developmental and operational tests and between operational tests and field performance. Knowledge of the ways in which developmental and operational tests are unreliable predictors of field performance has great value for reliability growth estimation, and could be useful both in linking developmental and operational test results and in providing information on how to design developmental and operational tests with greater operational realism. Meeker described the following possibility for addressing a linkage between developmental and operational testing.
Developmental tests are often accelerated, meaning that stresses are frequently increased in an effort to simulate the greater passage of time and greater use. To make accelerated testing informative for decisions concerning operational or field performance, a model (e.g., a degradation model) is used to relate accelerated test time to actual use time. This model must describe the effects of acceleration, the impact and distribution of environmental conditions, and the distribution of use rates in actual use of the system. (This type of model is often related to physics-of-failure models, discussed below.) A successful model of this form could be used to link developmental test data on system reliability to operational test and field performance. Meeker gave an example concerning the use of washing machines. Here the failure probability was expressed as a function of the number of cycles of use, and users were divided into categories based on their rate of use in cycles per unit time. Within these categories, the rate of use was assumed to be constant. Use of this assumption made it possible to translate the failure probability, initially expressed as a function of the number of cycles (which could be experimented on for high use) into a failure probability expressed as a function of time (see Meeker et al., 2002). Agreement between this mixture distribution and field performance can be used as a validation tool, validating, for example, that the percentages of people