Independent Component Subspace Method and Its Application to Fault Detection

ZHANG Muguang,SONG Zhihuan
2010-01-01
Abstract:To handle the modeling problem for high-dimension data,the independent component subspace method(ICSM) was proposed.As a new ensemble learning method,ICSM could overcome the main drawback of the random subspace method.It constructed subspaces according to independent components (ICs) contributions on different process variables.As a result,the modeling requirement of the ensemble learning method was satisfied,and its physical meaning was also well presented.Moreover,a new fault detection method named ICSM-PCA was also developed.Firstly,PCA monitoring models were build on different subspaces,then the weighted value of each model was computed based on T~2 and SPE statistics.Finally,two ensemble statistics could be built for monitoring industrial processes.A case study of the Tennessee-Eastman(TE) process illustrated that the proposed method showed good modeling performance and exhibited satisfactory fault detection ability.
What problem does this paper attempt to address?