ABSTRACT
Initial uncertainties exist commonly in observed degradation records, especially when non-periodic inspection policy is adopted. Previous research treated such initial uncertainties more as unwanted noises, and concentrated on how to control their influence on later prediction. However in this paper, we saw the initial uncertainties as a part of the stochastic degradation model. Supposing that the degradation process is a time-dependent Ornstein-Uhlenbeck (OU) process, we considered different hypotheses for initial uncertainties from a data-analysis and engineering view respectively, which were treated totally as 2 different models: one is with initial uncertainties and the other is with a constant start 0. The explanation in real applications for these 2 choices are clear. As the equipment’s degradation is recorded when it is “new”, so engineers think there should not be any degradation at all which appear in data as 0. However in real degradation data, we found that those “new” equipments may probably have initial degradations which appear as initial uncertainties subject to the measurement errors, imperfect production procedures etc. So we did some simulation tests to judge which hypothesis is better for the given degradation data and models’ performance on parameter estimation, evaluation results and prediction ability is compared when a real degradation data set is compared. It is concluded from the simulation results that the model considering initial degradation predicts first passage failure more conservatively than the one ignoring initial degradation based on the given degradation data. Also under Akaike information criterion and other criterions, the model considering initial degradation’s influence is better the one ignoring considering initial degradation.
