Determining the effectiveness of run-in: a case study in the analysis of repairable-system data

Run-in tests are often performed to prevent latent defects and other early life failures from reaching the customer. In the product under study, initial analysis of field data indicated that the original run-in length of 500 actuations was insufficient to surface all early failures. A selected group of machines was given an extended run-in of 5000 actuations and subsequent field data were collected. The effectiveness of the extended run-in over the original test as displayed by the field data was studied using several nonparametric methods (raw rate of occurrence of failure, Barlow's method, Nelson's mean cumulative number of failures), the parametric nonhomogeneous Poisson process (Crow's method), and a graphical exploratory analysis technique (correspondence analysis). Although one analysis would have been sufficient to answer the question of interest, several methods were evaluated to see if the conclusions were consistent and to better understand the advantages and disadvantages of each method. Each of the methods considered yields the same conclusion: the initial rate of occurrence of failures for the extended run-in group is less than that for the control group.<
>