Remaining useful life prediction based on self-attention mechanism -sequential variational autoencoder: From a semi-supervised perspective

Remaining useful life (RUL) prediction is a crucial task in maintaining operational safety and dependability of complex systems. Considering that conventional data-driven RUL prediction approaches demand massive labeled samples for supervised model training, it has notable limitations in making full use of unlabeled degradation data. Furthermore, current deep networks require in-depth research work in terms of interpretability and uncertainty. In this sense, an RUL prediction approach with the self-attention mechanism-sequential variational autoencoder (SAM-SVAE) is proposed from a semi-supervised perspective. Specifically, considering a dynamic serialization modeling, this paper designs a self-attention mechanism network to focus on key parts of the input time window. On this basis, this paper explores the correspondence between a Bayesian deep probability generation network and the state space model in control theory, which approximates the RUL prediction’s density function for uncertainty assessment. Moreover, this paper proposes an SAM-SVAE from a semi-supervised perspective, which can learn valuable feature representations from a large amount of unlabeled degradation data, from which the interpretability is provided through the analyze of latent space. Experimental validation of the presented SAM-SVAE utilizes the aircraft turbofan engine dataset from NASA Prediction Center.