Predicting Imprecise Failure Rates from Similar Components: A Case Study Using Neural Networks and Gaussian Procedures

Reliability prediction methods for early design stages suffer from the lack of empirical failure data. However, expert knowledge and data from similar in-service components are information sources ready available. In this work, an expert elicitation procedure for failure rate prediction is presented. In contrast to direct elicitation procedures such as FMEA, relations to in-service components are elicited. Neural networks and Gaussian processes are used to capture relations such as ”similarity” and ”wear out” to predict failure rate bounds and prediction uncertainty of new components. Both approaches are evaluated on a test set with raw failure rats. Results indicate that the method is a promising alternative to direct elicitation procedures.