Research And Application Of Spectral Clustering Algorithm Based On Manifold Distance Kernel

In recent years, Spectral clustering algorithms in pattern recognition to obtain a widerange of applications and become one of the most popular clustering analysis methods.Spectral clustering algorithm based on the spectral graph theory, The spectral clusteringalgorithm can overcome the drawback of traditional clustering algorithm on the convex graphcompared with traditional clustering methods.So that it can be converged to global optimalsolution.It constructs a simplified data space making use of the eigenvectors after spectraldecomposition that not onlt reduces the dimension of data but also gives clearer distributionof data in the subspace.In order to better reflect the relationship between the point and pointfor the original data, this paper presents a combination of manifold distance and spectralclustering and obtain a spectral clustering algorithm based on manifold distance kernel.In this thesis, a lot of research has been done for spectral clustering algorithm and itsapplication, which can be summered as follows:1.This paper introduces clustering analysis and spectral clustering algorithm in detail,and also describes basic knowledge of the graph, matrix repesentation, degree matrix, andlaplacian matrix, then makes a systematic exposition of spectral graph theory, rules of graphpartition and achievement of spectral clustering algorithm.2.In the standard spectral clustering algorithm, In connection with the similarity metric isbased on Euclidean distance, So a novel spectral clustering algorithm based on manifolddistance kernel is presented in this paper which can fully reflect the data clusteringcomplexity of the space distribution. It can exploit the inherent structure information of thedatasets and reflect the local and global consistency, Experimental results show that comparedwith traditional clustering and popular spectral clustering algorithms, this algorithm canachieve better clustering effect on a challenging artificial datasets and UCI public databases.Therefore, its improves the algorithm’s clustering performance finally.3.This paper regards the improved spectral clustering algorithm as a way ofunder-sampling A novel under-sampling unbalanced dataset SVM algorithm based onspectrum cluster of manifold distance kernel is presented, so the SVM classification performance under unbalanced dataset is improved. Finally use MDKSC-SVM algorithm onthe roller bearing failure detection dataset for testing, in the experiments, the proposedapproach is compared with other data-preprocess methods for unbalanced datasetclassification, the experimental results demonstrate that this algorithm has better detectionperformance.in rolling bearings fault detection.