Predictive planning of maintenance windows reduces the risk of unwanted production or operational downtimes and helps to keep machines, vessels, or any system in optimal condition. The quality of such a data-driven model for the prediction of remaining useful lifetime is largely determined by the data used to train it. Training data with qualitative information, such as labeled data, is extremely rare, so classical similarity models cannot be applied. Instead, degradation models extrapolate future conditions from historical behaviour by regression. Research offers numerous methods for predicting the remaining useful lifetime by degradation regression. However, the implementation of existing approaches poses significant challenges to users due to a lack of comparability and best practices. This paper provides a general approach for composing existing process steps such as health stage classification, frequency analysis, feature extraction, or regression models for the estimation of degradation. To challenge effectiveness and relations between the steps, we run several experiments in two comprehensive case studies, one from manufacturing and one from dry-bulk shipping. We conclude with recommendations for composing a data-driven degradation estimation process.
展开▼
