传统木结构复杂损伤的图神经网络诊断方法

Graph neural network-based diagnosis of complex damages in traditional timber structures

关键词:  传统木结构  结构健康监测  损伤识别算法  图神经网络

Key words:  traditional timber structure  structural health monitoring  damage identification algorithm  graph neural network

摘要:

针对传统木结构损伤识别算法在参数空间完备性、高维数据利用效率及局部整体损伤联合诊断能力上的不足,提出基于图神经网络（GNN）的多模态数据融合损伤识别方法。首先,以结构节点加速度响应以及材性参数为输入,融合传感器拓扑关系构建图结构数据,实现节点间损伤特征的交互传递与协同识别。其次,提出联合识别算法,由局部损伤识别网络（LCGCN-LDI）与整体材性劣化网络（GCN-MPI）组成,分别对节点进行损伤识别及对材性参数进行修正。最后,试验验证联合识别算法的结构损伤综合识别精度达94.7%,局部损伤准确率为97.6%,自振频率预测误差由传统模型的28.9%降至9.4%。结果表明,联合识别算法对传统木结构复杂损伤的识别效果优于传统算法,具备较强的精度与鲁棒性。

Abstract:

To address the limitations of traditional timber structure damage identification algorithms in parameter space completeness, utilization efficiency of high-dimensional data, and integrated local-global damage diagnosis capabilities, this paper proposes a multi-modal data fusion damage identification method based on a graph neural network (GNN). First, with the acceleration response of structural nodes and material parameters as inputs, the graph structure data is constructed by fusion of sensor topological relations to realize the interactive transmission and collaborative identification of damage features between nodes. Second, this paper proposes a joint recognition algorithm, which is composed of a local damage identification network (LCGCN-LDI) and a global material deterioration network (GCN-MPI), to identify the damage of nodes and correct the material property parameters respectively. Finally, experimental validation shows that the joint identification algorithm achieves an overall structural damage recognition accuracy of 94.7%, and a local damage detection accuracy of 97.6%; the prediction error of natural vibration frequency is reduced from 28.9% to 9.4%. The results show that the joint recognition algorithm outperforms traditional algorithms in identifying complex damage in traditional timber structures, exhibiting high accuracy and robustness.