A variational autoencoder based lightweight physics-informed neural network for remaining useful life prediction

Abstract Remaining useful life (RUL) prediction, as a pivotal technology in prognostics and health management, plays a critical role in proactive fault prediction and maintenance planning. Although data-driven methods have achieved remarkable progress in RUL prediction, their practical industrial application is hindered by high model complexity and poor interpretability. While existing physics-informed neural network (PINN) approaches can address these limitations, their neglect of computational quantity and uncertainty quantification in prediction results undermines their effectiveness for predictive maintenance. In this paper, we propose a variational autoencoder (VAE) based lightweight PINN for RUL prediction. The self-attention mechanism is introduced into a VAE to effectively extract degradation features and approximate the nonlinear dynamical equations of degradation systems as prior physical knowledge. Subsequently, the self-attention assisted VAE-PINN captures the underlying physical mechanisms during degradation and predict RUL. Validated on a benchmark dataset, the proposed method demonstrates superior prediction performance and improved interpretability with significantly fewer parameters compared to state-of-the-art approaches.