Bayesian Domain-Adversarial Regression Adaptation: A New Self-Driven Remaining Useful Life Prediction Method Across Different Machines With Uncertainty Quantification

Remaining useful life (RUL) prediction has shown substantial potentials in predicative maintenance for various kinds of machinery. In practical applications, however, the effect of RUL prediction would be much reduced due to lack of run-to-failure monitoring data of target machine, since a malfunctional component will be replaced before falling into a complete failure. It is obligated to perform RUL prediction by utilizing sequentially collected monitoring data, with the prognostic model dynamically updated. Subsequently, the other challenge is to quantify the degradation uncertainty in online scenario for more trustworthy prediction results. To address these concerns, a new self-driven RUL prediction method, named Bayesian domain-adversarial regression adaptation (BDRA), is proposed. BDRA consists of a Bayesian pretraining network and a self-driven updating mechanism. Running on domain-adversarial training framework, a Bayesian pretraining network is first established using available online data and offline degradation data. Through employing variational inference (VI) to approximate the posterior distribution of the regressor’s model parameters, domain-invariant temporal feature representation can be extracted. A new self-driven updating mechanism is then built on the online data chunks that sequentially arrive. Initialized by the pretrained feature representation, a monotonicity-aware self-supervised information is explored in form of core tensor. By integrating such self-supervised information as well as the pseudolabels generated from the pretraining network, adaptive RUL prediction, with a confidence interval (CI) for the prediction results, can be obtained. Taking rolling bearing as research object, this article conducts cross-condition and cross-machine experiments on two widely used bearing datasets, achieving state-of-the-art performance in terms of predictive accuracy and reliability.