in the various categories of use rate do not change over time. Divergences from these assumptions and identification of remedies are the subject of ongoing research.
Once these and other uses of field performance data have been institutionalized with the accompanying benefits, the quality of such data is likely to improve. One important application in which industrial field performance data have recently been improving in quality is sensors that can collect the entire history of use of, say, an automobile, including stresses, speeds of use, temperature, etc. That information can be linked to information on system reliability or performance to support much richer statistical reliability modeling. Use of such sensors could be valuable in operational testing for similar reasons.1 Several companies are undertaking efforts to save additional data in their warranty databases so the data can be used not only for financial purposes, but also for reliability assessment and estimation. Such efforts represent a cultural change. A hurdle is that development or expansion of such a database sometimes requires innovative funding approaches.
In his discussion of the papers by Sen and Gaver, Paul Ellner addressed the complication involved in reliability growth modeling of translating reliability estimates from developmental test to predictions of reliability in the field from operational test. At present, analysts may use a reliability growth model to extrapolate from developmental test results to operational test results. This approach can be severely biased, producing overly optimistic reliability estimates since the failure modes can be substantially different in the two situations (actually three—developmental test, operational test, and field performance). Efforts to perform this translation face the following challenges: (1) determining the (approximate) relationship between failure modes that occur jointly in developmental and operational environments, and (2) identifying a function expressing the probability of failure in operational test as a function of the probability of failure in operational test due to failure modes not present in developmental test. This translation probably cannot be done at the system level; it must be carried out at the com-