基于STGCN模型的大坝变形监测缺失数据处理OA

Aiming at the problem of missing data in dam deformation monitoring data and the limitation of traditional interpolation methods in capturing complex spatio-temporal correlation,this paper proposed a new method of missing data processing based on spatio-temporal graph convolutional network(STGCN).This method first abstracted the dam monitoring system into a topology diagram,where the nodes represent physical monitoring points.The edge structure construction of the graph comprehensively considered the physical proximity between the measuring points(such as the vertical line and the collimation line layout)and the data-driven correlation based on the Pearson correlation coefficient to accurately characterize the synergistic deformation relationship within the structure.Secondly,a multi-dimensional feature vector integrating dynamic and static information was constructed for each node.The vector covered deformation,key environmental loads(upstream water level and temperature),time coding reflecting periodicity,and static attributes such as node type and long-term historical statistics,which provided comprehensive information input for the model.On the model architecture,the temporal convolutional network(TCN)was used to capture the long-term temporal dependence of the deformation sequence,and the graph convolutional network(GCN)was used to propagate and aggregate information on the graph structure,so as to realize the collaborative learning of the spatio-temporal dynamic evolution pattern of the deformation field.In order to verify the performance of the model,this paper took the long-term measured data of 32 measuring points of Wuxi River Pier Dam as a case study.Compared with benchmark models such as multiple chain equation interpolation(MICE)and Kriging method,the results show that the STGCN model has significant advantages in various evaluation indexes such as determination coefficient(R2),mean absolute error(MAE),and root mean square error(RMSE).The model not only performs well on the positive vertical measuring points with strong regularity but also shows strong high-precision interpolation ability and generalization performance on the sight line measuring points with more complex environmental impact.In addition,the ablation test results clearly confirm that the introduction of a GCN module that characterizes spatial correlation is the key to improving the interpolation accuracy of the model.The research shows that the STGCN model proposed in this paper can effectively deal with the problem of missing dam deformation monitoring data.The interpolation results have high accuracy and are more in line with the physical deformation law of the structure.It can provide a reliable data basis for the subsequent development of high-precision deformation prediction,structural health diagnosis,and digital twin system construction.