Search Results

Fatigue life regression models with constant and non-constant variance are evaluated and compared with a Random Fatigue Limit Model and a Probit model to estimate the fatigue strength and S-N relationship from fatigue test data. The Maximum Likelihood method is used to estimate parameters of the above models. Emphasis also is given to assessing the variation in fatigue strength and S-N relationships estimated from different models. Two different test data sets are selected for the evaluations: (1) a full S-N test data set with a large range of stresses and (2) a staircase test data set with a characteristically more narrow range of stresses. Model adequacies are evaluated from residuals and Anderson Darling measures of their fits. The Random Fatigue Limit Model is observed to best fit the test data, but its large number of parameters are under constrained with staircase test data and require care to get good results.

Complete data sets (i.e., when all components within a given sample have explicit failure mileage) are a rarity in automotive field data analysis. More typically, only a few components of the sample would fail and the respective failure mileage would be recorded. In order to correctly estimate the reliability function, however, one would need to know the (censoring) mileage on the non-failed components. The paper discusses a procedure to estimate the censoring mileage and, ultimately, the reliability function for a component of interest. The paper further argues that a similar procedure can be applied to the total time on test estimation in the reliability growth analysis.

The paper discusses some statistical aspects of warranty repair predictions in automotive industry. The existing Renewal Process approach to model a single component repairable system is reviewed, and its limitations in terms of applicable TTF distributions are discussed. The Superimposed Renewal Process approach to model a multi component repairable system is then considered and its limitations in terms of the error of the superimposed aggregation are pointed out. Finally, a Monte Carlo approach is presented and its advantages over the conventional methods are discussed.