Remaining useful life prediction with limited run-to-failure data: A Bayesian ensemble approach combining mode-dependent RVM and similarity.

Accurate prediction of remaining useful life (RUL) is crucial for predictive maintenance of industrial systems. Although data-driven RUL prediction methods have received considerable attention, they typically require massive run-to-failure (R2F) data which is often unavailable in practice. If not properly addressed, training with a limited number of R2F trajectories not only leads to large errors in RUL prediction, but also causes difficulty in quantifying the prediction uncertainty. To address the above challenge, this paper proposes a Bayesian ensemble RUL prediction method that combines mode-dependent relevance vector machine (RVM) and trajectory similarity. Firstly, the proposed approach clusters historical R2F trajectories of unequal lengths into different degradation modes, and constructs RVM and similarity based predictions with improved accuracy by using mode-dependent libraries of kernel functions and similar trajectories. Secondly, the proposed Bayesian ensemble scheme fuses the RVM and similarity based predictions, and quantifies the associated prediction uncertainty even though the number of historical R2F trajectories are limited. In two case studies involving bearings and batteries, using only 11 and 16 R2F trajectories as training data, respectively, the proposed method reduces the mean absolute percentage error of RUL prediction by more than 20% compared to three existing methods. • Investigate RUL prediction using a limited number of run-to-failure trajectories. • Perform mode-dependent prediction to address different degradation modes. • Propose a Bayesian ensemble approach fusing RVM and similarity based predictions. • Generate confidence intervals for RUL predictions without using massive data. • Achieve smaller prediction errors with lower variance compared to existing methods. [ABSTRACT FROM AUTHOR]