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

Structural damage identification is a critical technique for ensuring the safety and reliability of large-scale civil structures. However, existing purely data-driven approaches based on deep learning still suffer from limited physical interpretability and generalization capability. To address this limitation, a physics-informed graph neural network framework, namely the Temporal Power Flow Graph Network (TPF-GNet), is proposed. The principal innovation of TPF-GNet is the Temporal Power Flow Propagation (TPFP) module, which explicitly characterizes energy transmission in structural dynamics and embeds dynamic power flow into the message-passing process of graph neural networks. TPF-GNet utilizes multi-sensor acceleration responses around a target node as input, reconstructs the target sensor's acceleration time history by simulating the energy flow process, and enables unsupervised damage detection and localization through reconstruction errors. Numerical simulations and scaled benchmark frame tests demonstrate that TPF-GNet surpasses conventional GNN and LSTM baseline models in both accuracy and physical interpretability. The results demonstrate that incorporating dynamic power flow markedly improves the model's ability to capture structural state changes induced by stiffness degradation or local damage. This study establishes a physics-constrained paradigm for structural health monitoring, particularly well-suited to engineering applications without damage labels.