Model selection for degradation-based Bayesian reliability analysis

• Investigate the reliability prediction model selection methods using multiple criteria.
• Predictor selection based on likelihood-based statistical metrics.
• Quantify within- and between-sample variations using random effect degradation models.
• Simulation study to compare model-based methods with non-parametric reliability analysis.

Traditional life testing for highly reliable products could become problematic in terms of an organization's fast product release and other competitive advantages such as market share and new technology introduction. Such issues result in the wide application of utilizing degradation data for reliability analysis and prediction for highly reliable products. The multi-unit degradation data can be either aggregated for reliability inference or modeled by mixed effect degradation models assuming given degradation path functions. This paper focuses on degradation model selection by exploring various combinations of fixed effects and random effects included in the degradation path models. The paper provides a systematic way for model selection using both statistical and empirical criteria. The model selection process is demonstrated by modeling the crack length data as a function of the number of cycles of a metal material under multiple hierarchical linear and log-linear models. These random effect models are able to capture both within-component variations due to repeated measurements over time and between-component variations due to unit-to-unit variations of the multiple sampled materials. Simulated data based on the selected hierarchical linear models are used to estimate the survival function, and the results are compared with those estimated from the non-parametric method using Kaplan–Meier estimation.