Hyperbolic Hypergraph Neural Networks for Hierarchical Fault Diagnosis in Rotating Machinery

Area of Science:

• Machine Learning
• Artificial Intelligence
• Mechanical Engineering

Background:

• Intelligent fault diagnosis is crucial for industrial safety and reliability.
• Existing hypergraph neural networks (HGNNs) often fail to capture hierarchical fault structures due to their Euclidean space operation.
• Modeling complex, hierarchical fault-response relationships requires advanced methods.

Purpose of the Study:

• To propose a novel Hyperbolic Hypergraph Neural Network (H²GNN) for intelligent fault diagnosis.
• To address the limitations of Euclidean-based HGNNs in representing hierarchical fault structures.
• To improve the accuracy and reliability of rotating machinery fault diagnosis.

Main Methods:

• Developed H²GNN integrating hyperbolic geometry with HGNNs.
• Constructed fault-response-aware hyperedges using diagnostic views of vibration signals.
• Employed Poincaré ball model for message passing and introduced adaptive curvature mechanisms.

Main Results:

• Achieved 99.87% accuracy on the CWRU bearing dataset and 99.75% on the MFPT bearing dataset.
• Outperformed six competing methods including CNN, GCN, and various HGNN variants.
• Ablation studies confirmed the significant contribution of hyperbolic geometry and adaptive curvature.

Conclusions:

• H²GNN effectively models hierarchical fault-response structures in rotating machinery.
• The proposed framework offers superior performance for intelligent fault diagnosis.
• Hyperbolic geometry provides a more suitable representation for complex diagnostic tasks.