Feature-Enhanced Multisource Subdomain Adaptation on Robust Remaining Useful Life Prediction

Abstract

In prognostic and health management (PHM), the remaining useful life (RUL) prediction is one of the key tasks. However, a complete run-to-failure record (with label) is not always available on some specific machines, for example, the new machine. To address the issue, the new machine may refer to other machines of the same or similar type (even old machines with labels) for developing the prediction model. Transfer learning, particularly domain adaptation, is one methodology used to transfer the knowledge gained from the source domain to the target domain. This study proposes feature-enhanced multi-source subdomain adaptation (FEMSA) to predict the RUL. FEMSA learns the domain-invariant features and characterizes the similarity by redefining the multiple source domains; that is, we handle the cross-domain generalization by reformulating the multiple operating conditions. In the experimental study, two datasets are applied to validate the proposed FEMSA, and the result shows that the FEMSA can provide more robust RUL prediction over time and thus improve the PHM system.