How data-driven approaches can increase the accuracy of LED reliability predictions

This paper proposes a methodology on the classification of LED degradation time-series to a suitable fitting model based on standard LED aging test data. Therefor a data set of about 2000 LED degradation curves was analyzed regarding their projected lifetime of various extrapolation models. The outcomes were compared to the standard exponential projection model. Results indicate a deviation of the projected L70 lifetime of more than 10.000 h in over 50 % of the cases. A classification on the data set was performed with respect to each functions best fitting model as its class label. Three types of classifiers were evaluated: Time Series Forest, Learning Shapelets and k-Nearest Neighbor. Results indicate that the Time Series Forest classifier predicts in over 75 % of the cases the most accurate fitting model in terms of accuracy and computation time. Likewise, k-Nearest Neighbor yields good accuracies with downside of low computational speed. Learning Shapelets on the other hand performed poorly supposedly due to the low time resolution of the given degradation time-series.