Fault detection in the Tennessee Eastman benchmark process using dynamic principal components analysis based on decorrelated residuals (DPCA-DR)

Abstract

Current multivariate control charts for monitoring large scale industrial processes are typically based on latent variable models, such as principal component analysis (PCA) or its dynamic counterpart when variables present auto-correlation (DPCA). In fact, it is usually considered that, under such conditions, DPCA is capable to effectively deal with both the cross- and auto-correlated nature of data. However, it can easily be verified that the resulting monitoring statistics (T2 and Q, also referred by SPE) still present significant auto-correlation. To handle this issue, a set of multivariate statistics based on DPCA and on the generation of decorrelated residuals were developed, that present low auto-correlation levels, and therefore are better positioned to implement SPC in a more consistent and stable way (DPCA-DR). The monitoring performance of these statistics was compared with that from other alternative methodologies for the well-known Tennessee Eastman process benchmark. From this study, we conclude that the proposed statistics had the highest detection rates on 19 out of the 21 faults, and are statistically superior to their PCA and DPCA counterparts. DPCA-DR statistics also presented lower auto-correlation, which simplifies their implementation and improves their reliability.