Making Reliability Estimates when Zero Failures are Seen in Laboratory Aging

Abstract

To estimate the probability of failure for a deployed population, when the laboratory aging of a sample certified population produces zero failures, an appropriate model is required. The widely used exponential (constant hazard rate) model, and a declining-hazard-rate Weibull model are contrasted. On balance the use of the Weibull model seems preferred because: (i) It is consistent with laboratory aging studies on many types of electronic components; (ii) It is less sensitive to the choices for parameters (e.g., thermal activation energy) whose values are not known accurately; and, (iii) It may yield more conservative (i.e., larger) values for the relevant hazard rate than the exponential model. The unavoidable uncertainties connected with the use of either model are discussed. An example is given of a plausible, but in fact mistaken, use of the exponential model that leads to disastrously incorrect reliability predictions.