Improvement Of Spectral Clustering Algorithm And Its Application In Bauxite Flotation Conditions Recognition

Abstract:As an efficient clustering algorithm, the clustering quality of spectral clustering is not affect by the class shape of the cluster sample point. And regardless of the sample space is convex or not, it can obtain the global optimum. However, it makes spectral clustering has certain limitations while the clustering process is sensitive with the scale parameter and need to set the number of clusters artificial in advance. In order to get the scale parameter and optimal number of clusters adaptively, this paper focuses on the spectral clustering algorithm and for the simulation of bauxite flotation conditions recognition based on machine vision.(1) Analyzed the basic idea of spectral clustering algorithm and its related theory, than described the algorithm steps of classic spectral clustering.(2) Through constructing a new similarity function using the dot density of data points’neighborhood to adjust the similarity between the data points, to make it more consistent the similar relationship between data points of the actual clusters, solved the problem that Spectral clustering algorithm is sensitive to the scale parameter when constructing the similarity matrix. Meanwhile, the scale parameter is obtained adaptively by neighbors’ distance when calculating the similarity, constitute the adaptive spectral clustering algorithm based on density adjustments to overcome the difficulty of set the scale parameter artificially.(3) As traditional spectral clustering need to determine the number of clustering in advance, this paper based on the relative positional relationship in feature space of the data points which is from the actual classes in the dataset is proposed, by calculating the angle between the mapping points in the feature space of all data points from data sets when setting the different clusters number to determine the optimal number of clusters, and an improved adaptive spectral clustering algorithm is proposed by combining with the adaptive spectral clustering based on density adjustment.(4) Taking the bauxite flotation process as the research object, the paper use the improved adaptive spectral clustering which is proposed to do simulation experiments with the image features of the flotation bubbles collected by the flotation process conditions recognition system based on machine vision. Experimental results show that the improved adaptive spectral clustering algorithm can accurately identify the conditions of the bauxite flotation process based on the froth image features and verify the effectiveness of the proposed algorithm.