A Big Data Architecture for Fault Prognostics of Electronic Devices: Application to Power MOSFETs

Abstract

This paper deals with the problem faced performing prognostics of electronic devices using a data-driven approach to generate degradation models for predicting their remaining useful life. To be able to generate good models, a lot of experimental data are required. Moreover, the high frequency sampling required for electronic devices implies that huge amounts of experimental data must be efficiently stored, transformed, and analyzed in the prediction models. The first contribution of this paper is the proposal of a Big Data architecture that can be used for a generic prognostics approach of electronic devices. To illustrate the proposal, the dataset for power MOSFET prognostics developed at the NASA Prognostics Center of Excellence is used. This paper carefully illustrates the analysis, extraction, and transformation stages required to obtain the data for the estimation of the degradation models. An additional contribution of this paper is to study scalable methods to perform such estimation. Instead of using typical approaches such as extended Kalman filters, particle filters, or relevance vector machines to perform the estimation, we propose to use much simpler techniques (such as least squares or horizontal average) to allow a scalable implementation in a Big Data (distributed and parallelized) platform. After applying our approach to the MOSFETs dataset, we have shown that the obtained results are competitive when compared with more complex techniques.