GMVE: Graph-Mamba variational encoder for interpretable remaining useful life prediction with uncertainty quantification

Predicting the remaining useful life (RUL) of industrial systems is essential for effective prognostics and health management. However, current deep learning approaches often fail to adequately model complex spatiotemporal dependencies in multi-sensor data and lack the interpretability crucial for maintenance decision-making. This study presents a graph-Mamba variational encoder (GMVE) that addresses these limitations through three key innovations: a graph-edge attention mechanism that captures dynamic inter-sensor relationships, the integration of these relationships into Mamba's selective state space model for efficient temporal modeling, and a variational framework that enables both uncertainty quantification and interpretable representation of system degradation. The GMVE maps degradation patterns into a probabilistic latent space where mean and variance parameters enable accurate RUL predictions with uncertainty estimates. Experiments conducted on benchmark datasets reveal that the GMVE outperforms state-of-the-art methods while offering valuable insights into equipment health evolution. The proposed approach effectively unifies spatiotemporal dependency modeling with uncertainty-aware interpretable predictions, thereby advancing prognostics for complex industrial systems.