Traditional data-driven graph convolutional networks (GCN) often lack physical interpretability and suffer from performance drops during domain shifts. To overcome these issues in rolling bearing fault diagnosis, we introduce a physical feature-fused attention GCN (PhyAttGCN) that directly embeds physical priors into the feature learning process. First, we extracted multi-dimensional physical features, including fault orders and time-domain statistics, then a dynamic gated fusion unit is designed to adaptively screen and fuse these physical features with data-driven features. Second, a dual-view graph construction strategy is proposed. Rather than relying on a single topology, the graph is constructed from two independent perspectives: a data view based on feature similarities and a physical view grounded in mechanical principles. An attention mechanism then fuses these two perspectives to form a stable cross-domain graph structure. By feeding this combined graph into a multi-receptive field GCN(MRF-GCN), the model gathers topological details across different scales simultaneously to ensure accurate fault classification. Tests on the CWRU and MFS datasets show that our framework performs much better across different operating conditions and datasets while maintaining clear physical meaning. This method provides a highly reliable solution for bearing fault diagnosis under complex and severe conditions.

cross-domain fault diagnosis; graph convolutional network; physical feature fusion; dual-view graph