301 / 2025-06-19 11:23:20

Physics Guided Dynamic Weighting Graph Neural Network for Remaining Useful Life Prediction of Rolling Bearings

Rolling bearings,graph neural network,physical information,graph attention mechanism,remaining useful life prediction

Graph neural network (GNN) has demonstrated the ability to learn feature representations from graph data and has been effectively applied to the task of predicting the remaining useful life (RUL) of machines. However, existing methods fail to fully incorporate the physical prior knowledge inherent in the bearing degradation process, thereby hindering models from accurately capturing the coupling relationship between complex degradation dynamics and failure mechanisms. To address this limitation, this paper introduces a GNN architecture guided by physical information, embedding the constitutive relationship between input feature evolution and physical degradation laws into the GNN structure via regularization constraints. Specifically, we propose a physics-informed loss function that minimizes the RUL prediction error during the bearing degradation process while adhering to fundamental physical principles. By dynamically adjusting weights according to the trend changes in monitoring data, the loss function of model is further refined, thereby enhancing its training process. Moreover, an attention mechanism is integrated to optimize the spectral graph convolution operation, mitigating the adverse effects of ambiguous features during continuous spectral graph convolution. Experimental results on two bearing datasets demonstrate that the proposed method successfully incorporates degradation-related physical information, leading to significant improvements in RUL prediction performance compared to existing approaches.