Assessing multi-output Gaussian process regression for modeling of non-monotonic degradation trends of light emitting diodes in storage

Abstract

Light emitting diodes (LEDs) exhibit different degradation physics under different environmental conditions of humidity, temperature and electrical loading, leading to complex degradation models - a common behavior with several other electronic devices. While most researches focus on degradation under active use, degradation models in storage are often not well established. Large fleet storage of components, in the absence of a degradation model, requires laborious continuous inspections despite the preservation under similar environmental conditions. Leveraging on training data from other LEDs within the fleet, stored under similar conditions, this study investigates the utility of multi-output Gaussian Process Regression (MOGPR) with limited test data, to model the complex degradation curve of LEDs in storage, as a proxy for electronic components. We further explore the choice of detrending means and training data sets, to enhance the prediction of degradation curves and residual storage life (RSL). Additional training data sets are observed to give diminishing returns for prediction accuracy.