Statistical Considerations in Predicting Operating Life from Accelerated Life Tests

A major problem in assuring the reliability of devices constructed using new materials or processes is that, initially, reliability can only be estimated through the results of accelerated life tests. The models constructed from these life tests are based partially on physics and partially on statistical modelling. The final models are transformations either of the time scale or degradation rate (1/time). Unfortunately for most physically plausible models, these transformations are highly nonlinear. The subject of this talk is a review of the methods that my colleagues at AT&T Bell Laboratories and I have developed over the past few years to estimate those transformations, and a discussion of some of the problems we are currently working on. The methods include estimation procedures for both failure time and continuous degradation data. The problems include approximate methods for quick calculations, experiment design for identifying with the analysis of continuous degradation data.