A Bayesian spatio-temporal degradation model based on ST-NaBGCN for aero-engine remaining useful life prediction and uncertainty quantification

Abstract With the advent of big data era, deep learning (DL) techniques have made remarkable progress and shown promising results in remaining useful life (RUL) prediction. However, existing DL models often suffer from overconfidence in their predictions. The absence of reliable uncertainty quantification not only limits the trustworthiness of the outputs but also increases the tendency to overfit. Moreover, while graph convolutional networks (GCNs) are commonly employed to capture spatial dependencies among sensor signals, their use of a greedy neighborhood aggregation mechanism leads to over-smoothing and limits their ability to effectively fuse and extract informative features. To address these challenges, this paper proposes a novel Bayesian spatio-temporal (ST) degradation model, named ST neighborhood-adaptive Bayesian GCN (ST-NaBGCN), for accurate RUL prediction with uncertainty quantification. The model integrates neighborhood-adaptive graph convolutions and gated recurrent unit to enhance ST representation and incorporates a Bayesian neural network to provide reliable uncertainty estimates. Two benchmark experiments on the CMAPSS and N-CMAPSS datasets are performed to evaluate ST-NaBGCN’s ability to represent uncertain spatio-temporal dependencies and thereby support degradation feature extraction and uncertainty quantification. Experimental results demonstrate that ST-NaBGCN not only achieves superior predictive accuracy compared to existing methods but also provides reliable uncertainty estimates.