Predicting hardware failure using machine learning

Abstract

The Weibull distribution has historically been the Reliability Engineer's best tool for describing the probability of failures over time [1]. While this technique is very accurate at describing failure distributions for large populations of components, it works very poorly at predicting the time until failure of an individual component. The mean time until failure is often used to predict times until failure of individual components, but this value may vary greatly with actual times until failure. With the advent of machine learning techniques, the ability to learn from past behavior in order to predict future behavior makes it possible to predict an individual component's time until failure much more accurately. In this paper, we explore the predictive abilities of a machine learning technique to improve upon our ability to predict individual component times until failure in advance of actual failure. Once failure is predicted, an impending problem can be fixed before it actually occurs. This paper brings to light a machine learning approach for predicting individual component times until failure that we will show is far more accurate than the traditional MTBF approach. The algorithm built was able to monitor the health of 14 hardware samples and notify us of an impending failure well ahead of actual failure, providing adequate time to fix the problem before actual failure occurred.