Few-shot remaining useful life prediction based on Bayesian meta-learning with predictive uncertainty calibration

Deep learning-based methods have been developed and widely used for predicting remaining useful life (RUL), a critical task in predictive maintenance aimed at minimizing machine downtime and optimizing maintenance schedules. These methods typically rely on substantial degradation data. However, in some real-world scenarios, degradation data are extremely limited. Meta-learning, a prominent few-shot learning method, leverages degradation data from auxiliary tasks to facilitate predictions in the target tasks. While meta-learning has improved prediction accuracy, there is inevitably uncertainty in predictions. To provide decision-makers with more reliable information, accurately quantifying predictive uncertainty is crucial. However, in meta-learning, both quantifying and calibrating predictive uncertainty are challenging due to data scarcity. This paper proposes a few-shot learning method for RUL prediction, named Bayesian model-agnostic meta-learning with predictive uncertainty calibration (BMLPUC), incorporating an uncertainty calibration term into the objective function for the model training. Additionally, the meta-training is optimized using two adaptive hyperparameters according to the model performance. The effectiveness of the proposed method is validated using two bearing datasets. Superior prediction accuracies and more accurately calibrated predictive uncertainty compared to the baseline and three other state-of-the-art methods are achieved by the proposed method.