Physics-Aware Mamba With Dynamic Hypergraph for Rolling Bearing RUL Prediction

Abstract

Remaining useful life (RUL) prediction for rolling bearings is important for achieving predictive maintenance of equipment. However, during actual equipment operation, variations in working conditions often lead to unstable RUL prediction results. To overcome these limitations, a physics-aware Mamba prediction model integrated with dynamic hypergraphs (DyG-Mamba) is proposed in this article. In particular, the framework follows a hierarchical alignment-calibration paradigm. First, a multisource feature extractor is pretrained through domain adversarial training, thereby enhancing the capability of the model to extract domain-invariant features from vibration signals. Subsequently, a physics-aware bidirectional Mamba module is designed. This module uses differential variables of key degradation indicators as physical priors to adaptively adjust the discretization step size of the state-space model, thereby accurately capturing the nonuniform evolution rates across different degradation stages. Finally, a dynamic hypergraph mixture mechanism is proposed, which dynamically constructs hyperedges containing multiple nodes in the feature space. By capturing higher order correlations among samples, it aggregates global manifold information of similar physical states, achieving cross-sample and cross-domain feature alignment and correction. Verification is conducted using the XJTU-SY bearing dataset. The results indicate that the proposed method outperforms several state-of-the-art methods across several evaluation metrics.