EXPONENTIALLY WEIGHTED MOVING PRINCIPAL COMPONENTS-ANALYSIS AND PROJECTIONS TO LATENT STRUCTURES

For stable (non-dynamic) chemical processes characterized by multivari ate data, principal components analysis (PCA) and projections to laten t structures (PLS) have recently been shown to provide useful monitori ng schemes. In this work, PCA and PLS are generalized to dynamically u pdated models for modelling processes with memory and drift. These mod els are based on exponentially weighted observations, and are formulat ed as multivariate generalizations of the exponentially weighted movin g average (EWMA). Principles and estimation algorithms for EWM-PCA and EWM-PLS are presented, and predictive control schemes based on these models are discussed.