Filter-based feature selection for prognostics incorporating cross correlations and failure thresholds

<jats:p>Historical condition monitoring data from technical systems can be utilized to develop data-driven models for predicting the remaining useful life (RUL) of similar systems, whereas the Health Index (HI) often is a crucial component. The development of robust and accurate models requires meaningful features that reflect the system’s degradation process, enabling an accurate prediction of the system's HI. Traditionally, the identification of those is supported by one of various feature ranking methods. In literature, feature interdependencies and their transferability across various similar systems are not sufficiently considered in feature selection, exacerbating the challenge of HI prediction posed by the scarcity of data and system diversity in real-world applications. This work addresses this gaps by demonstrating how filter-based feature selection, incorporating failure thresholds and cross correlations, enhances feature selection leading to improved HI prediction. The proposed methodology is applied to a novel dataset* obtained from run-to-failure experiments on geared motors conducted as part of this study, which presents the aforementioned challenges. It is revealed that classical feature selection, consisting of feature ranking only, leaves potential untapped, which is utilized by the proposed selection methodology. It is shown that the proposed feature selection methodology leads to the best result with a RMSE of 0.14 in predicting the HI of a constructive different gearbox, while the features, determined by classical feature selection, lead to a RMSE of 0.19 at best.</jats:p>