Multivariate Process Monitoring Using Classical Multidimensional Scaling and Procrustes Analysis

This paper presents a new process monitoring framework using multidimensional scaling. The traditional method of multivariate process monitoring is generally base on principal component analysis (PCA) and is carried out by monitoring the fault detection parameters Hotelling's T2 and squared prediction errors (SPE). Both indexes are derived directly from multivariate scores in the observation sample configurations. This conventional system was found inappropriately used especially in monitoring highly nonlinear multivariate processes leading to a great number of principal components being selected. Alternatively, classical multidimensional scaling (CMDS) is another technique which can be used in compressing multivariate data by using dissimilarity measures for process monitoring. The proposed process monitoring system is developed based on variable relationships and the dissimilarity measures in terms of variable profiles are used in projecting the multivariate scores. A new monitoring index, which is the resultant vector length different between the new and the normal variable profiles, is introduced. Procrustes analysis (PA) is implemented for on-line process monitoring through a moving-window mechanism. The proposed monitoring method is demonstrated on a simulated continuous stirred tank reactor (CSTR) with recycle system. The results show that the proposed system was efficient as well as effective in detecting various abrupt and incipient faults compared to the linear PCA-based scheme.