Unsupervised Kernel Regression Modeling Approach for RUL Prediction

Abstract

Recently, Prognostics and Health Management (PHM) has gained attention from the industrial world since it aims at increasing safety and reliability while reducing the maintenance cost by providing a useful prediction about the Remaining Useful Life (RUL) of critical components/system. In this paper, an Instance-Based Learning (IBL) approach is proposed for RUL prediction. Instances correspond to trajectories representing run-to-failure data of a component. These trajectories are modeled using Unsupervised Kernel Regression (UKR). A historical database is used to learn a UKR model for each training unit. These models fuse the run-tofailure data into a single feature that evolves over time and hence allow the construction of a library of instances. When
unseen sensory data arrive, the learned UKR models are used to construct the test degradation trajectories. RUL is deduced by comparing the test degradation trajectory to the library of instance. Only the most similar train instances are kept for RUL prediction. The proposed approach was tested and compared to approaches that apply linear regression and PCA to model the library of instances. Results highlight the benefit of using UK compared to other approaches.