Bias Suppression Framework for Detrending Mean of Multi-Output Gaussian Process Regression in LED Remaining Storage Life Prognosis

Advancements in storage prognosis tend to be limited by the inherent challenge to collect sufficient significant degradation data over an extensive period. Using only sparse fleet data, multi-output Gaussian process regression (MOGPR) is one of the few techniques that offers a practical data driven approach to model non-monotonic degradation profiles with low mean absolute percentage error (MAPE). This accuracy in the
storage
prognosis context is, however, sensitive to the choice of the
detrending mean
. Working with light emitting diodes (LED) sparse lumen degradation data under storage conditions in this study, the MAPE is observed to be highly correlated to the
detrending bias
– the difference between the detrending mean and the test mean. We explore various approaches to suppress this bias and advocate a generic framework for fleet storage prognosis. The approaches include
detrending
using (A) static training data mean, (B) dynamic observed test data mean, (C) static bounded training data set pairs, (D) dynamic weighted mean of unbounded training data set pairs and (E) moving average of weighted mean of unbounded training data set pairs. Our analysis shows that the moving average approach (Method E) of computing weighted mean of unbounded training data set pairs results in the most stable detrending mean to suppress detrending bias and helps achieve an MAPE lower than 1%.