화학공학소재연구정보센터
Industrial & Engineering Chemistry Research, Vol.43, No.22, 7036-7048, 2004
Multidimensional visualization of principal component scores for process historical data analysis
Principal component analysis (PCA) is an effective approach for projecting process operation onto a reduced dimensional space. If the first two principal components (PCs) can capture an adequate amount of the variance of the data, the two-dimensional plot of the scores of the two PCs provides a visual representation of the process operating envelopes. However, because of the difficulty in visualizing more than three dimensions in Euclidean space, which requires the coordinates to be perpendicular to each other, it has not been possible to visually represent the operating envelopes when more than three PCs have to be considered. This paper presents an approach for multidimensional visualization of multiple PCs using a technique called parallel coordinates for the purpose of process monitoring. The effectiveness of the approach is demonstrated by applying it to a database corresponding to 527 days of operation of a wastewater treatment plant and by comparing its performance against those of the well-established multivariate statistical process control (MSPC) and a conceptual clustering algorithm that were previously applied to the same database. It was found that both PCA visualization and the MSPC T-2 chart identified the same 17 days as "clearly abnormal" and another 8 days as "likely abnormal". The conceptual clustering identified 14 of the 17 clearly abnormal days.