Fusion of Manifold Learning and Spectral Clustering Algorithmwith Applications to Fault Diagnosis

Large amount of multivariate data in many areas of science raises the problem of data analysis and visualization. Focusing on high dimensional and nonlinear data analysis, an improved manifold learning algorithm is introduced, then a new approach is proposed by combining adaptive local linear embedding (ALLE) and recursively applying normalized cut algorithm (RANCA). A novel adaptive local linear embedding algorithm is employed for nonlinear dimension reduction of original dataset. The recursively applying normalized cut algorithm is used for clustering of low dimensional data. The simulation results on three UCI standard datasets show that the new algorithm maps high-dimensional data into low-dimensional intrinsic space, and perfectly solves the problem of higher dependence on the structure of datasets in the traditional methods. Thus classification accuracy and robustness of spectral clustering algorithm are remarkably improved. The experiment results on Tennessee-Eastman process (TEP) also demonstrate the feasibility and effectiveness in fault pattern recognition.