The use of machine learning for electrical component end-of-life predictions

Abstract

Diminishing Manufacturing Sources and Material Shortages (DMSMS), also referred to as obsolescence, is a sector of product sustainment that is receiving more attention as certain technologies continue to have longer and longer system life cycles. Much of the research today points towards a need for better electrical component end-of-life (EOL) predictors. A small-scale case study was performed to explore the use of machine learning for obsolescence forecasting of flash memory chips. The Random Forest classification algorithm was able to predict the Active vs. Obsolete status in both the training data and the test data with an OOB error rate of 10.87%. The Random Forest regression algorithm was able to predict an obsolescence date of an obsolete component on average 0.75 years after the actual discontinuation of the component and 1.08 years for active components. The regression analysis had an overall error rate of 0.53%. This study demonstrates opportunities and challenges for using machine learning as a future DMSMS forecasting tool.