Research On Dam Abnormal Data Detection Based On Clustering Ensemble

With the development of national water resources,a large number of dams with international leading level have been built in China.Reservoir dams produce many social and economic benefits,such as flood control,water supply,irrigation,power generation and tourism.However,the huge potential energy generated by flood storage also poses a great threat to the safety of people’s lives and property in the downstream.The main purpose of dam safety monitoring is to master the operation characteristics of the dam and the change trend of each monitoring measurement.With the rapid development of modern monitoring technology,the dam monitoring points are more comprehensive,and the amount of monitoring data is huge.If the monitoring data are processed and analyzed one by one,it will need a lot of manpower and time,which is not conducive to grasp the dam operation status in time.In order to evaluate the engineering safety,how to quickly and accurately extract useful information from the massive monitoring database has become a very valuable research topic.In this paper,the actual monitoring data of an earth rock dam is taken as the sample,and the cluster integration method model is used to analyze the monitoring point information and select the abnormal monitoring point data,so as to further study the deformation law of the dam face.(1)In order to make full and timely use of the abundant monitoring information,this paper introduces a semi supervised learning method based on constraints on the basis of traditional spectral clustering algorithm for the data distribution of dam safety monitoring data.Based on the prior knowledge of samples(constraints between samples),the constraint matrix of sample points is constructed,and the constraint information ml and CL are used appropriately to improve the clustering effect to a certain extent.(2)With the increase of constraint pair,the clustering accuracy of constrained spectral clustering algorithm will increase,but the constrained spectral clustering algorithm is not sensitive to density,that is,when the local density of samples is relatively large,the clustering effect of traditional constrained spectral clustering algorithm is not very good.In order to overcome this deficiency,on the basis of the above research,the idea of shared nearest neighbor and natural neighbor is used for reference Density can adjust the similarity between data points.This paper presents an improved constrained spectral clustering method for dam face monitoring.In this method,the natural eigenvalues generated by natural neighbors are combined with shared neighbors,and the similarity matrix is redefined,so that it is no longer sensitive to parameters,and the cluster structure of dam panel monitoring data set is effectively explored.(3)Considering the distribution characteristics of abnormal data points in historical dam safety monitoring,the selection of initial clustering center points is optimized according to the density characteristics of data points.Based on the K-means++ algorithm which is based on the maximum distance of clustering center and the principle of density priority,the basic clustering generator is designed.At the same time,the integrated analysis strategy is introduced,and the improved constrained spectral clustering algorithm is used as the consistency function to integrate the basic clustering results.The dam anomaly data detection model based on cluster integration is designed.The dam panel deformation monitoring data are clustered and analyzed,and then the optimal classification is carried out The key abnormal points of dam face are determined.Detailed analysis of the key abnormal points can not only reduce the workload,but also grasp the real operation status of the dam in time.This method can avoid the algorithm falling into the local optimal solution,reduce the number of iterations,and improve the quality of clustering and the efficiency of anomaly detection.Taking the dam panel monitoring data as the research object,an experimental data set is formed,which further verifies that the anomaly data detection model based on clustering ensemble can effectively deal with the clustering problem of multi-scale data sets,and is superior to the traditional anomaly detection method based on clustering in detection accuracy and efficiency.