Physics-Guided Adaptive Graph Transformer for Multi-Modal Bearing Fault Diagnosis Under Variable Working Conditions

Multi-sensor fusion provides richer information for bearing fault diagnosis. However, under variable working conditions, the coupling relationships among signals from different sensors exhibit significant non-stationarity and directionality, posing challenges for modeling and practical deployment. Existing methods often rely on fixed or symmetric graph structures or construct correlation relationships entirely based on data-driven approaches; this makes balancing physical consistency, robustness, and computational efficiency difficult. To address these issues, we propose a Physics-guided Adaptive Graph Transformer Network (AGTN) for multi-modal bearing fault diagnosis under variable working conditions. More specifically, we offer innovative improvements across three aspects. Firstly, we introduce domain knowledge priors into the graph structure learning process to adaptively construct sparse and asymmetric dynamic graph structures that capture physically meaningful directional dependencies among different sensor signals. Secondly, we combine a graph-aware transformer to jointly model the temporal features and structural correlations of multi-source signals. Finally, we further introduce a hierarchical subgraph training strategy that significantly reduces memory usage and training time while ensuring diagnostic performance. Experimental results on a self-built multi-condition bearing dataset show that AGTN achieves an average diagnostic accuracy of 99.42% under the same distribution conditions and demonstrates good generalization and robustness, e.g., variable speed and load and sensor failure. In particular, when using only 25% of the nodes for training, the model can still maintain a diagnostic accuracy of 97.9%, while reducing the peak memory usage to about 19% of that of full-graph training. The above results validate the effectiveness of the proposed method under complex industrial conditions, as well as its practical application potential in resource-constrained scenarios.