Bayesian large-kernel attention network for bearing remaining useful life prediction and uncertainty quantification

Abstract

Attention network-based remaining useful life (RUL) prediction methods have achieved distinguished performance due to the ability of adaptive feature selection. However, existing attention networks fail to balance between the computational efficiency and the long-range correlations as well as channel adaptability. Moreover, these attention networks are unable to reason about the uncertainty in RUL prediction. To tackle these issues, a Bayesian large-kernel attention network (BLKAN) is proposed for bearing RUL prediction and uncertainty quantification. BLKAN enables uncertainty quantification, long-range correlations and channel adaptability in attention mechanism to effectively extract degradation features to facilitate RUL prediction accuracy. Thereafter, large kernel Bayesian convolutions, that are used to generate attention weights in BLKAN, are decomposed into three simple components to reduce the parameters and computational cost. At last, variational inference is introduced to inference probability distributions of the parameters of BLKAN and learn uncertainty-aware attention. Experimental results on two bearing datasets show that BLKAN not only achieves uncertainty quantification in RUL prediction but also consistently outperforms the baseline comparison methods. Visualization of attention weights reveals the causal correlations between the degradation patterns and the features emphasized by attention. The proposed method provides a novel uncertainty-aware attention network-based framework for trustworthy RUL prediction.