Research On Machine Learning Models For Health Diagnosis Of Spatial Deformation Behavior Of Super-high Arch Dams

Structural health monitoring is one of the important ways used for dam safety management.In order to evaluate the real working behavior of super-high arch dams,this study is carried out from three perspectives:parameter inversion,monitoring model and risk rate analysis,which can provide technical support for the scientific decision-making of the dam safety management department.Firstly,on the basis of fully considering the spatial association of multi-monitoring points of the dam body,an incremental distance-based hierarchical clustering method is proposed to divide the measured displacement field of an arch dam into different zones.Aiming at the abnormal deformation behavior of the Jinping-I arch dam that the displacement continues increasing in the water level stable stage,based on the HST model,the hysteretic hydraulic component is introduced,and the HEST model is established to separate the viscoelastic total hydraulic displacement component.The multi-monitoring points zonal inversion of dam concrete viscoelastic parameters is realized with the utilization of the ELM-IPSO.Then,based on the optimal viscoelastic parameters of the inversion results,radial displacement of a concrete arch dam under hydraulic load can be calculated by the finite element method,and then the HHST models are established by using machine learning technologicies of the SVM and RVM,respectively.At the same time,in order to solve the problem of over-fitting caused by the single objective of only minimizing the fitting MSE in traditional optimization modeling method,a double objective function,in which the fitting MSE is as small as possible and the shape similarity of multiple displacement time series is as large as possible,is proposed and used to optimize the parameters of machine learning models.In order to further improve the performance of machine learning models,a double objective optimized machine learning combination prediction model is established with the MLR,NN,ELM,SVM and RVM as sub-models.According to the 3S control principles used for structural health monitoring of dams,the distribution law of residuals between predicted and measured dam displacements is analyzed,and the real-time risk rate of a single moitoring point is defined on this basis.Finally,the Copula function,which can reflect the correlation of multiple monitoring points,is used to establish the real-time joint risk rate model of an crch dam.Research indicates:(1)The zonal inversion results of multiple moitoring points can take the deformation spatial integrity of an arch dam and the difference of boundary constraints into account,thus it can better reflect the real working behavior of high arch dams.The inverted instantaneous elastic moduli of the four zones of the Jinping-I arch dam are 44.8,48.8,30.8 and41.0 GPa,respectively.(2)The combination weight of the two objectives,the fitting MSE and SSI,has a significant impact on the performance of machine learning models.The average prediction accuracies of the double objective-optimized SVM models,which uses the HST and HHST causal factors,are improved by 50.8% and 47.4%,and the average overfitting indexes are reduced by 44.3% and 70.9%,respectively.The average decrease of the RMSE and MAE of the RVM model is 31.2% and 24.8%,respectively.The multi-kernel function can further improve the prediction performance of the SVM and RVM model.The prediction confidence bandwidth of the RVM model is significantly smaller than that of the MLR model,with an average decrease of 75.1%.(3)The joint risk rate of an arch dam has a negative exponential function correlation with the number of used monitoring points.For the Jinping-I arch dam studied in this paper,when the number of monitoring points exceeds 10,the joint risk rate tends to be stable.The measured displacement data-based high risk period of arch dams mainly occurs in the stage when the change rate of the upstream reservoir water level changes sharply.Therefore,the change rate of water level should be slowed down as much as possible at the beginning and end of reservoir regulation.