Prognostic Health Management for LED with Missing Data: Multi-task Gaussian Process Regression Approach

Light emitting diodes (LEDs) are increasingly used in a number of applications such as display backlighting, signaling, general illumination, medical services etc. They have high efficiency, environmental resilience, high reliability and long lifetime. Normally, one has to resort to using sequential updating methods such as the particle filtering approach for effective prognosis and lumen maintenance life prediction of LEDs because of their long lifetime and reliability. However, the performance of the particle filter severely depends on the selection of the initial parameters. Moreover, it cannot deal with situations where there is missing data due to intermittent sensor failure or malfunction. The multi-task Gaussian process regression (MTGP) can model multiple correlated multivariate time series simultaneously. Thus, it can be used to learn the correlation between multiple degradation trends (from different similar units tested) even though they might be sampled at different frequencies and have training sets available only for different time spans. This paper presents the multi output (task) - Gaussian process regression as a solution for prognosis of lumen maintenance life of LED devices in the event of "missing" data. We analyze the performance of the MTGP on the LED data extracted from previously published studies and demonstrate that it can give good prediction results even when there are missing data.