A Process Monitoring Method Based on Global-Local Structure Analysis in Principal Component Reconstruction Space

Abstract

The performance of projection-based process monitoring methods is prone to be affected by noise in data. By constructing latent variables that have large contribution to variance, principal component analysis (PCA) can separate the main information from noise. Thus, it is widely used in data-based process monitoring and signal denoising. Global-local structure analysis (GLSA) can extract both global information and local structure information by combining PCA and local preserving projection. However, GLSA cannot avoid the possibility to model noise information. In this paper, a novel method is proposed which applies GLSA in the data reconstruction space of principal components. This method can not only extract global variance information and local structure information but also avoid the possibility of noise modeling. Simulations based on data from a numerical example and TE process demonstrate that the proposed method is superior to GLSA for fault detection.