Field-Reliability Predictions based on Statistical System Life Cycle Models

Reliability measures the ability of a system to provide its intended level of service. It is influenced by many factors throughout a system life-cycle. A detailed understanding of their impact often remains elusive since these factors cannot be studied independently. Formulating reliability studies as a Bayesian regression problem allows assessment of their impact simultaneously and to identify a predictive model of reliability metrics. The proposed method is applied to currently operational particle accelerator equipment at CERN. Relevant metrics were gathered by combining data from various organizational databases. To obtain predictive models, different supervised machine learning algorithms are applied and compared in terms of their prediction error and reliability. Results show that the identified models accurately predict the mean-time-between-failure of devices – an important reliability metric for repairable systems - and reveal factors which lead to an increased dependability. These results provide valuable inputs for early development stages of highly dependable equipment for future particle accelerators.