Application of multi-output Gaussian process regression for remaining useful life prediction of light emitting diodes

Abstract

Light-emitting diodes (LEDs) are the preferred technology today when it comes to lighting both for indoor and outdoor applications, predominantly due to their high efficiency, environmental resilience and prolonged lifetime. Given their widespread use, there is a need to quickly qualify them and accurately predict the reliability of these devices. Due to their inherently long operational life, most LED reliability studies involve the use of degradation tests and application of filter-based prognostic techniques for dynamic update of degradation model parameters and estimation of the remaining useful life (RUL). Although they are in general very effective, the main drawback is the need for a specific state-space model that describes the degradation. In many cases, LED degradation trends are affected by a multitude of unknown factors such as unidentified failure modes, varying operational conditions, process and measurement variance, and environmental fluctuations. These variable factors that are hard to control tend to complicate the selection of a suitable state-space model and in some cases; there may not be a single model that could be used for the entire lifespan of the device. If the degradation patterns of LEDs under test deviate from the state space models, the resulting predictions will be inaccurate. This paper introduces a prognostics-based qualification method using a multi-output Gaussian process regression (MO-GPR) and applies it to RUL prediction of high-power LED devices. The main idea here is to use MO-GPR to learn the correlation between similar degradation patterns from multiple similar components under test and thereby, bypass the need for a specific state space model using available data of past units tested to failure.