Accurate prediction of the remaining useful life (RUL) of aircraft engines is crucial for ensuring flight safety and optimizing maintenance strategies. However, traditional data-driven methods typically yield point estimation, failing to quantify uncertainties arising from data noise and model limitations. This study proposes an attention-based ensemble method with partial-transfer Bayesian deep learning (Att-ensembled PT-BDL) for uncertainty quantification in aircraft engine RUL prediction. The proposed method transfers weights and biases from existing point estimation deep learning models as prior knowledge to the mean values of weights and biases in Bayesian deep learning models, freezing these parameters during training to reduce trainable parameters and enhance computational efficiency. An ensemble framework, enhanced by an attention mechanism, integrates multiple models to improve prediction accuracy and uncertainty quantification performance. A case study is conducted to demonstrate the effectiveness of the proposed method with a dataset of the PHM data challenge. The experiment results show that the proposed Att-ensembled PT-BDL method can achieve a better prediction accuracy and uncertainty quantification performance in terms of root mean square error (RMSE), prediction interval coverage probability (PICP) and prediction interval normalized average width (PINAW).
Zeng et al. (Tue,) studied this question.