Non-causal data-driven monitoring of the process correlation structure: A comparison study with new methods

Current approaches for monitoring the process correlation structure lag significantly behind the effectiveness already achieved on the detection of changes in the mean levels of process variables. We demonstrate that this is true, even for well-known methodologies such as MSPC-PCA and related approaches. On the other hand, data-driven process monitoring approaches are typically non-causal and based on the marginal covariance between process variables. We also show that such global measure of association is unable, by design, to effectively discern changes in the local correlation structure of the system and propose, for the first time, the explicit use of partial correlations in process monitoring. As a second contribution, we introduce the use of sensitivity enhancing data transformations (SET) with the ability to maximize the detection ability of all monitoring procedures based on (partial or marginal) correlation, and show how they can be constructed. Results confirm the added-value of the proposed monitoring scheme.