A Novel Approach for Analyzing Censored Lifetime Data from Designed Reliability Improvement Experiments

Design of experiment has been widely used to improve product reliability and a number of case studies have been documented in the literature. Primary features of reliability improvement experiments are the presence of censored data and small sample size. These make data analysis very challenging. Various methods have been developed to analyze such data. However, some research gaps still exist, e.g., how to select an appropriate performance measure as response, how to estimate the response from the experiment data, and how to identify significant factors and possible interactions. This paper aims to address these issues through reanalyzing the industrial thermostats experiment data. We compare the robustness and appropriateness of the B10 life and negative quadratic mean (which is equivalent to the Taguchi signal-to-noise ratio) as the performance measures and conclude that the B10 life is more appropriate. A robust method that involves both non-parametric and parametric approaches is developed to estimate the B10 life. A correlation-coefficient-based and computationally simple approach is also developed to identify significant two-factor interactions. Applying the proposed measure and methods to the thermostats data yields the reasonable results. Thus, the paper provides useful methods for analyzing the reliability improvement experiment data.