A reliability approach for prediction and management of part obsolescence for improved system health

Abstract

Accurate prediction of part obsolescence is critical to maintaining system health, especially for the long-lived systems typical in aerospace and naval domains. While there are methods that predict an expected date of obsolescence, a numerical likelihood of obsolescence can be useful. This work describes a Weibull-based conditional probability method for the prediction of part-level obsolescence risk. Several considerations inherent to the problem environment and using a probabilistic method to estimate risk are discussed and addressed, including accounting for changing product life, using dynamic binning and Weibull regression; sample bias, through data cleaning; and small datasets with potentially highly censored data, using a modified synthetic minority oversampling technique (SMOTE) that can sample both the minority and majority classes. Development of an approximate measure of uncertainty of obsolescence is also presented. Use of the method is demonstrated with a multiplexer dataset and shows the feasibility of the approach.