A prognostic driven predictive maintenance framework based on Bayesian deep learning

Abstract

Recent years have witnessed prominent advances in predictive maintenance (PdM) for complex industrial systems. However, the existing PdM literature predominately separates two inter-related stages—prognostics and maintenance decision making—and either studies remaining useful life (RUL) prognostics without considering maintenance issues or optimizes maintenance plans based on given/assumed prognostic information. In this paper, we propose a prognostic driven dynamic PdM framework by integrating the two stages. In the prognostic stage, we characterize the latent structure between degradation features and RULs through a Bayesian deep learning model. By doing so, the framework is capable of generating a predictive RUL distribution that can well describe prognostic uncertainties. In the maintenance decision-making stage, we dynamically update maintenance and spare-part ordering decisions with the latest predictive RUL information, while satisfying operational constraints. The advantage of the proposed PdM framework is validated by comparison with several benchmark polices, based on the famous C-MAPSS turbofan engine data set.