Research on structural damage identification based on temporal power flow graph network.

Area of Science:

• Civil Engineering
• Structural Health Monitoring
• Artificial Intelligence

Background:

• Data-driven deep learning methods for structural damage identification lack physical interpretability and generalization.
• Existing approaches struggle with unsupervised learning and require labeled damage data.

Purpose of the Study:

• To develop a physics-informed graph neural network framework, TPF-GNet, for enhanced structural damage identification.
• To improve the physical interpretability and generalization capability of deep learning models in structural health monitoring.
• To enable unsupervised damage detection and localization without requiring labeled damage data.

Main Methods:

• Proposed the Temporal Power Flow Graph Network (TPF-GNet) framework.
• Introduced the Temporal Power Flow Propagation (TPFP) module to embed dynamic power flow into graph neural networks.
• Utilized multi-sensor acceleration responses for unsupervised damage detection and localization via reconstruction errors.

Main Results:

• TPF-GNet demonstrated superior accuracy and physical interpretability compared to conventional GNN and LSTM models.
• The TPFP module effectively captured structural state changes caused by stiffness degradation or local damage.
• Validated through numerical simulations and scaled benchmark frame tests.

Conclusions:

• TPF-GNet establishes a physics-constrained paradigm for structural health monitoring.
• The framework offers improved performance and interpretability for engineering applications, especially in unsupervised scenarios.
• Incorporating dynamic power flow is crucial for accurately assessing structural integrity.