A Multiscale Method for Predicting the Remaining Useful Life of Bearings Using Physics-Informed Hypergraph Wavelet Attention Network
Abstract
Accurate prediction of bearing remaining useful life (RUL) is crucial for developing optimized mechanical equipment maintenance strategies and safeguarding the reliability of system operation. Recently, for complex higher order correlations hidden in monitoring data, hypergraph-based methods have demonstrated superior modeling and feature representation capabilities. However, these existing methods fail to incorporate physical information constraints into model training and the construction of topological structures, resulting in a lack of model interpretability and limiting the improvement of model evaluation accuracy. To overcome this limitation, we propose a physics-informed hypergraph wavelets attention network (PI-HGWAT) for evaluating the RUL of bearings. First, a physical constraint mechanism is introduced to guide the network in learning parameters that satisfy wavelet mathematical properties, thereby ensuring multiscale decoupling of bearing degradation features while adhering to physical characteristics.Second, a hypergraph topology is explicitly constructed based on the wavelet packet transform decomposition tree to model complex hierarchical relationships, thereby endowing the topological structure with physical significance. Then, a tri-stage hypergraph attention network is developed to achieve dynamic aggregation of multiscale degradation features in bearings. Finally, a bidirectional gated recurrent unit is employed to capture temporal dependencies within degradation trajectories. Comprehensive experiments conducted on two public datasets demonstrate the notable effectiveness and superior performance of PI-HGWAT.