Data-driven condition-based maintenance optimization given limited data

Unexpected failures of operating systems can result in severe consequences and huge economic losses. To prevent them, preventive maintenance based on condition data can be performed. Existing studies either rely on the assumption of a known deterioration process or an abundance of data. However, in practice, it is unlikely that the deterioration process is known, and data is often limited (to a few runs-to-failure), especially for new systems. This paper presents a fully data-driven approach for condition-based maintenance (CBM) optimization that is especially useful in situations with limited data. The approach uses penalized logistic regression to estimate the failure probability as a function of the deterioration level and allows any deterioration level to be selected as the preventive maintenance threshold, also those that have not been observed in the past. Numerical results indicate that the preventive maintenance thresholds resulting from our proposed approach closely approach the optimal values.