Failure-time prediction

Abstract

This paper considers methods for predicting the number of future failures of items of any kind manufactured on consecutive days, when the survival times of failed items are given. Of particular interest is the case where failures do not occur uniformly, but tend to cluster in certain stretches of adjacent days. This feature of the data suggests that in order to get a good model the failure data ought to be pooled over segments of adjacent days. An algorithm based on dynamic programming has been developed which finds the optimal number of segments as well as the segments themselves. Failure time for each segment is modeled either by an exponential or a Weibull distribution, which ever is better as judged by the minimum description length (MDL) principle to minimize the stochastic complexity of all the observed failure times. The same criterion is used to define the optimality of a collection of segments. The technique can be extended to include other usual model classes. Predictions of the number of failures in any future time interval can be made using the optimal segments and their models. © 1998 Elsevier Science B.V. All rights reserved.