Research On Dynamic Warning And Prediction Method For Deformation Of Concrete Dam Based On Spatiotemporal Features Learning

The developing trend prediction and abnormal behavior of dam deformation are the main tasks of dam safety monitoring and controlling.However,a large amount of multi-source data has been accumulated during long-term continuous monitoring,and the dam deformation and its mapping relationship with related factors show obvious spatiotemporal differentiation,due to the influence of the spatial difference and time variability of service environment and material properties.Which pose a higher requirement for the efficiency and accuracy of dam deformation prediction and early warning,putting forward the following issues urgent to be solved: First,under the background of massive multi-source information,how to systematically interpret the spatiotemporal differentiations and the dynamic effects of its driving factors.Second,in the analysis of deformation behavior,how to achieve reasonable zoning of dam deformation.Third,in the prediction of deformation,how to accurately establish the mapping rules between the multi-point displacement and the environmental factors under the consideration of the deformation spatiotemporal differentiation.Finally,in the early warning of deformation safety,how to effectively mining and construct the indepth mapping between the multi-point deformation,environmental factors and dam service status.Therefore,focusing on the above problems,it is of great practical significance to carry out the research on deformation safety prediction and early warning methods using spatiotemporal feature learning.In our work,a real concrete arch dam is taken as an example to develop in-depth research and the main results are as follows:(1)The methods for analyzing the spatiotemporal differentiation and dynamic effects of driving factors in the concrete dam deformation are proposed based on the Gaussian space distance based exploratory spatial data analysis(Gaussian-ESDA)algorithm and the hybrid grey dynamic incidence model(HGDIM).In view of the status that the current research on dam deformation behavior focuses on the average displacement characteristics of the dam in time or space dimension,and fails to reveal the spatiotemporal differentiation of deformation and the dynamic effect of its influencing factors,the analyzing methods of the spatiotemporal differentiation and dynamic effects of driving factors for the concrete dam deformation are developed based on Gaussian-ESDA and HGDIM.First,taking the Gaussian spatial distance measure to improve the way of weight calculation in ESDA algorithm,a GaussianESDA-based deformation spatiotemporal differentiation analysis method is proposed to counter the deficiency of distinguish the deformation correlation differences between non-equidistant measuring points in the conventional ESDA algorithm.Second,coupling the similarity measurement methods based on the curve shape and changing trend of data,the HGDIM-based dynamic effects analysis method of deformation driving factors is established to address the disability of the comprehensive measurement of the dynamic correlation between deformation and its influencing variables only by the similarity analysis of a single geometric feature in classical grey incidence model.The real case study shows that the Moran’s I index of deformation at each measuring point derived from the proposed method is greater than 2.58 at the significance level of 0.01,and compared with the traditional statistical model,the prediction accuracy of the method considering the dynamic effects of driving factors is improved by at least 18%,which can reveal the deformation spatiotemporal differentiation features and dynamic effects of its driving factors of dams in multidimensional and multi-scale manner.(2)The Gaussian-mixture-model Clustering by Fast Search of Density Peaks with Grey Wolf Density Entropy Optimization(GWODP-GMMC)based deformation behavior zoning method is proposed for the concrete dam.To counter the problem that it is difficult to accurately identify the irregular deformation spatiotemporal differentiation characteristics in the current deformation behavior zoning which based on single clustering algorithm with distance-similarity measure,resulting in the deviation of zoning results from reality,the GWODP-GMMCbased clustering zoning method for deformation of concrete dams is developed.In which,the information entropy and grey wolf optimization algorithm are introduced to solve the issue of the reasonable setting of truncation distance in the clustering by fast search of density peaks(DPC)algorithm,and to initialize the model parameters of Gaussian mixture model clustering(GMMC)algorithm.By the integration of fast search of cluster centers in the DPC and iterative updating of model parameters in GMMC,the disadvantage that a single clustering rule cannot fully capture and identify the characteristics of spatiotemporal differentiation is settled.The real case study shows that the results of the proposed hybrid clustering model can correctly reflect the spatial coordination and seasonal periodicity of dam deformation,and compared with the with the traditional single clustering algorithm,the average clustering accuracy is improved by 6%.(3)The multi-monitoring points deformation prediction model of concrete dam is established based on clustering ensemble learning under the influence of spatiotemporal differentiation characteristics.To deal with the problems that the current research on deformation prediction mostly focuses on the regression analysis of the displacement trend of single point,and the commonly used single-algorithm machine learning-based model(SAMLM)in dam deformation prediction that failed to fully capture the mixed nonlinear relationship between multi-point deformation and environmental factors under different spatiotemporal differentiation characteristics,a multi-monitoring points deformation prediction model based on the improved DP-GMM-clustering-based ensemble learning(IDP-GMMC-EL)method is established under the influence of spatiotemporal differentiation feature.First,the multi-output ensemble learning framework(MOELF)coupled with multiple training rules is constructed through the multi-output extension and integration of extreme learning machine(ELM)and support vector machine(SVM),which overcomes the deficiency that the SAMLM cannot completely learn the complex nonlinear mapping relationship between deformation and its influence factors on different parts of the dam.Furthermore,coupled the GWODP-GMMC-based zoning model with the MOELF,and embedded a synchronous optimization strategy based on grey wolf optimization improved by Levy flight and chaotic local search,the ensemble learning of multi-point deformation law in various spatiotemporal partition.The case analysis results show that the proposed model shares excellent synchronous prediction performance in multi-monitoring points deformation,which reduce the average prediction error by 35% on the basis of conventional statistical models.(4)The dynamic early warning model for concrete dam deformation is established based on deep robust kernel learning and cloud aggregation of spatiotemporal characteristics.To counter the deficiency that most of the deformation early warning methods use static probability statistics to analyze the deformation warning value,ignoring the evolution trend in spatiotemporal features of the in-depth nonlinearity between the multi-monitoring point deformation and environmental factors in space,as well as the random-fuzziness of deformation time-domain distribution,a dynamic early warning model for spatial deformation is proposed based on deep robust kernel learning and cloud aggregation of spatiotemporal features.First,embedding the proper orthogonal decomposition and the Correntropy-based noise constraint strategy into the deep kernel extreme learning machine,the disadvantages of the traditional shallow learning methods that are difficult to mine the deep mapping rules between the spatial principal components of deformation and environmental factors are overcome.Moreover,coupling the mutual transformation and fusion of random and fuzzy concepts in cloud model,and the detection and updating mechanism of concept drift theory,the reasonable formulation of deformation early warning indicators and reliable early warning of status for concrete dam under dynamic service environment can be realized.The real case study shows that by using the proposed early warning model,the prediction accuracy of the deformation spatial principal component is improved by 24%,and the obtained deformation warning interval with the advantage of dynamic updating,is narrower than those of the other methods,which reduces the risk of emergency underreporting.