On-line process monitoring using local measures of association. Part II: Design issues and fault diagnosis

• A new fault diagnosis scheme is proposed based on the use of partial correlations.
• A theoretical equivalency relationship is established between the EWMA constant and window size.
• An analytical expression is derived for the control limits in the off-line case.
• A Monte Carlo approach is proposed to compute the control limits in the on-line case.
• Partial correlations allow for a better isolation of faults' location.

Control charts based on partial correlations have proved to be an effective approach to detect fine deviations on the underlying structure of process data. The prompt detection of such faults is dependent on the proper selection of the control chart design parameters, namely their control limits, subgroup size (off-line case) and forgetting factor (on-line case). In this article, specific guidelines are provided to attain the desired detection power while maintaining the intended false alarm rate. A formal relationship that relates the on-line monitoring approach with the simpler off-line implementation is also derived. This relationship can then be used to design on-line control charts based on insights and results obtained with the more interpretable off-line version. A new fault diagnosis procedure is also introduced in order to take advantage of the partial correlations ability to remove the effects of faulty variables in the data, and thus obtain higher identification accuracy and decrease the total time invested in diagnosis activities and troubleshooting.