Power Bond graph-embedded causal flow-aware graph neural network for fault localization of electro-hydraulic systems

To achieve precise fault localization for electro-hydraulic systems of mechanical, a Causal Flow-aware Graph Neural Network (CFAGNN) is proposed, which integrates power bond graph topology with deep learning. A power bond graph model is constructed to characterize energy transfer mechanisms and clarify dynamic coupling relationships among subsystems, providing quantifiable physical constraints and edge-node attributes. To accommodate real-time operating conditions, learnable edge attributes are constructed with embedded dynamic correction parameters. These physical priors are embedded into the Causal Flow Attention (CFA) mechanism, where physical regularization and causal direction loss are utilized to constrain attention coefficients to conform to energy transfer laws, thereby enabling accurate tracking of fault propagation paths. Validation on a real engineering machinery dataset demonstrates that the proposed method outperforms state-of-the-art models. The results indicate that CFAGNN balances diagnostic accuracy and computational efficiency effectively, providing a reliable solution for the fault diagnosis of complex electro-hydraulic systems.