PMBCT: The Probabilistic Multiscale Bayesian Convolutional Transformer for Trustworthy Remaining Useful Life Prediction

Abstract

In industrial remaining useful life (RUL) prediction, the uncertainties can deteriorate generalization and cause low trustworthiness and accuracy of RUL prognostics results. To address this issue, a novel probabilistic multiscale Bayesian Convolutional Transformer (PMBCT) is proposed for trustworthy RUL prognostics with uncertainty quantification. Specifically, we design a Bayesian convolutional probsparse self-attention to integrate local context into global modeling and a multiscale representation learning mechanism to fuse scale-aware information, which can help the PMBCT to both globally and locally quantify uncertainty information and capture degradation features from diverse temporal scales. Moreover, to reduce the adverse effects induced by uncertainties on RUL prediction, we develop a Bayesian backpropagation training algorithm in which uncertainty information can be feedback to train the proposed model, improving its generalization. Finally, comprehensive RUL prediction experiments are carried out based on a bearings dataset for validating the effective and competitive performances of the PMBCT-based RUL prognostic approach.