Accelerating Optimal Test Planning With Artificial Neural Networks

In this paper, artificial neural networks for use in an algorithm for determining an optimal test plan are investigated. The neural networks serve to estimate the Probability of Test Success and the expected duration of a reliability test for a product with prior knowledge about the failure distribution. This estimation in negligible time enables the efficient pre-selection of promising test plans. Without such a pre-selection, a comparison of all possible test plans would lead to an unfeasibly large amount of simulation due to the high number of parameters. It is shown that this method is applicable for the most common reliability test types, including different kinds of censoring. Building on previous studies, several layouts of shallow neural networks are trained, tested and compared. It is shown that for all test types, a mean error on the Probability of Test Success of below 2.5 % can be reached. There is no systematic error to be found and the yielded neural networks are reproducible with high stability. As a result, fully connected two and three layer designs can be recommended for most test types. Their application allows a fast and therefore practical comparison of test plans made up from several common test types to find the one fitting best to the given product and reliability requirements.