A Fault Detection Framework Based on Diffusion Maps and Procrustes Analysis

Due to the wide use of distributed control systems, multivariate statistical process monitoring has drawn increasing attention. Dimension reduction of the high-dimensional data is quite fatal and vital for the process monitoring since the extracted latent variables represent the inherent characteristics of the process data. Traditional dimension reduction methods like PCA have been deeply studied in the area of process monitoring. Comparing to PCA, diffusion maps is nonlinear manifold learning method derived on the Markov process and robust to noise perturbations. Fault detection consists of off-line modeling and on-line monitoring. As diffusion maps could not get explicit map, Procrustes analysis is applied to facilitate the on-line monitoring. In this paper, we propose a new fault detection framework called Diffusion Maps and Procrustes Analysis (DM-PA). The performance of this new proposed fault detection framework is tested on the TE benchmark.