Multiscale Multivariate Statistical Process Control

Abstract

Dynamic process, both continuous and batch, are characterised by autocorrelated measurements which are allied to the effects of process dynamics and disturbances. The common multivariate statistical process control (MSPC) approaches have been to use principal component analysis (PCA) or projection to latent structures (PLS) to build a model that captures the simultaneous correlations amongst the variables, but that ignores the serial correlation in the data during normal operations. Under such conditions it is difficult to perform efficient fault detection and diagnosis. An alternative approach to account for the process dynamics in MSPC is to use multiresolution analysis (MRA) by way of wavelet decomposition. Here, the individual measurements are decomposed into different scales (or frequencies) and the signals in each decomposed scale are then used for MSP which provides an indirect way of handling process dynamics.