Yonatan Saadona, Noam Auslanderb, and F. Patrick McCluskeya
a CALCE/Department of Mechanical Engineering,
University of Maryland,
College Park, MD, USA
b The Wistar Institute,
Philadelphia
PA, USA
For more information about this article and related research, please contact Prof. Patrick McCluskey.
Abstract:
Accurate prediction of the remaining useful life (RUL) of a degrading component is crucial to prognostics and health management for electronic systems, to monitor conditions and avoid reaching failure while minimizing downtime. However, the shortage of sufficiently large run-to-failure datasets is a serious bottleneck impeding the performance of data-driven approaches, and in particular, those involving neural network architectures. Here, we develop a new data-driven prognostic method to predict the RUL using an ensemble of quantile-based Long Short-Term Memory (LSTM) neural networks, which represents the RUL prediction task to a set of simpler, binary classification problems that are amenable for prediction with LSTMs, even with limited data. We demonstrate that this approach obtains improved RUL estimation accuracy for power MOSFETs, especially with a small training dataset that is characterized by a wide range of the RUL.
This article is available online here and to CALCE Consortium Members for personal review.