Accelerated Life Test Modeling Using Median Rank Regression
Rhodes, Austin James
MetadataShow full item record
Accelerated life tests (ALT) are appealing to practitioners seeking to maximize information gleaned from reliability studies, while navigating resource constraints due to time and specimen costs. A popular approach to accelerated life testing is to design test regimes such that experimental specimens are exposed to variable stress levels across time. Such ALT experiments allow the practitioner to observe lifetime behavior across various stress levels and infer product life at use conditions using a greater number of failures than would otherwise be observed with a constant stress experiment. The downside to accelerated life tests, however, particularly for those that utilize non-constant stress levels across time on test, is that the corresponding lifetime models are largely dependent upon assumptions pertaining to variant stress. Although these assumptions drive inference at product use conditions, little to no statistical methods exist for assessing their validity. One popular assumption that is prevalent in both literature and practice is the cumulative exposure model which assumes that, at a given time on test, specimen life is solely driven by the integrated stress history and that current lifetime behavior is path independent of the stress trajectory. This dissertation challenges such black box ALT modeling procedures and focuses on the cumulative exposure model in particular. For a simple strep-stress accelerated life test, using two constant stress levels across time on test, we propose a four-parameter Weibull lifetime model that utilizes a threshold parameter to account for the stress transition. To circumvent regularity conditions imposed by maximum likelihood procedures, we use median rank regression to fit and assess our lifetime model. We improve the model fit using a novel incorporation of desirability functions and ultimately evaluate our proposed methods using an extensive simulation study. Finally, we provide an illustrative example to highlight the implementation of our method, comparing it to a corresponding Bayesian analysis.
- Doctoral Dissertations