Moving PCA for process fault detection - A performance and sensitivity study

Abstract

Conventional PCA has a major disadvantage of being less effective with time-varying and/or processes with multiple operation modes. In order to deal with the limitation, this thesis proposes Moving PCA (MPCA), which is based on updating scaling parameters (mean and standard deviation) from a moving window. Simulation results show that MPCA performs better than other approaches including conventional PCA, adaptive PCA, and Exponentially Weighted PCA in monitoring Tennessee Eastman Process (TEP) simulation and analyzing an industrial data set. In addition, MPCA robustness is investigated empirically by varying parameters including moving window size, number of principal components retained, and confidence limits. The results indicate that MPCA is more sensitive in analyzing the industrial data set than in monitoring TEP and moving window size seemed to be less critical than the other two parameters. Several monitoring indices including conventional statistics (T^2 and Q), combined QT and standardized Q index are also implemented in MPCA. The results demonstrate that T^2 and Q work equally well or better compared to the later two indices.