An Improved Spectral Clustering Algorithm Using Geodesic Distance

An improved spectral clustering algorithm is proposed to focus on the problem that the general clustering algorithms are invalid for reciprocating compressor fault data lying on complex manifold.A new affinity matrix is obtained.The geodesic distance replaces the traditional Euclidian distance to measure the similarity of data,and neighborhood-based density factor is used to identify and to remove noise points.Moreover,density-based local Euclidian distance adjustment is introduced into areas with small gap between manifolds.The proposed method is implemented on several artificial datasets and a real reciprocating compressor fault dataset.Experimental results show that the new algorithm can accomplish the clustering for data with noise and multi-scale character,especially when the manifolds have small gaps or crossover between each other.Its accuracy is 50.86% and 8.6% higher than those of k-means and MSCA respectively.