Physics-driven Bayesian long short-term memory network for machinery remaining useful life prediction with uncertainty estimation

Remaining useful life (RUL) prediction plays an essential role in machinery health management and reliability assessment. Although there are extensive researches on machinery RUL prediction, these methods remain several gaps in comprehending equipment degradation mechanisms and estimating prediction uncertainties. This paper develops a novel interpretable RUL prediction framework named Physics-driven Bayesian long short-term memory (PDBLSTM) network to address the above issues. First, the particle filter is adopted to infer hidden state information of the degraded machine's dynamic response. Subsequently, the physics data fusion is achieved by aggregating hidden state information and visible degradation features to enhance the input space of the developed BLSTM model for accurate RUL prediction and uncertainty estimation. Moreover, a novel customized physics embedding loss function is designed to constrain the network learning process to be consistent with generalized physics knowledge. The developed physics data fusion and customized physics embedding loss function improve the accuracy and interpretability of machine RUL prediction while reducing uncertainty. Finally, the effectiveness and superiority of the Physics-driven framework are evaluated using run-to-failure planetary gearbox datasets and rolling bearing datasets. Extensive experimental results demonstrate that the proposed PDBLSTM method achieves accurate RUL prediction and quantitative assessment of uncertainties in the degradation process.