Canonical variate analysis-based monitoring of process correlation structure using causal feature representation

• A new algorithm is proposed for monitoring changes in process structure.
• Algorithm combines canonical variate analysis with causal network analysis.
• Method has potential for detecting structural faults and identifying root causes.
• The effectiveness is demonstrated for single and multiple faults in case studies.
• The method performed better than correlation-based and variable-based methods.

Although the monitoring of process variables has been extensively studied, techniques for monitoring faults in the process correlation structures have not yet been fully investigated. The typical methods based on the covariance matrix of the process data for process monitoring have limited capability to effectively monitor underlying structural changes. This paper proposes a canonical variate analysis (CVA) approach based on the feature representation of causal dependency (CD) for the monitoring of faults associated with changes in process structures, which employs CD for pretreating the data and subsequently utilizes CVA for quantifying dissimilarity. Apart from the improved performance of capturing the underlying connective structure information, the utilization of the CD feature in the first step provides more application-dependent representations compared with the original data, as well as decreased degree of redundancy in the feature space by incorporating causal information. The effectiveness of the proposed CD-based approach for the monitoring of structural changes is demonstrated for both single faults and multiple faults in simulation studies of a networked system. In the simulation results, the CD-based method performs the best, followed by correlation-based and then variable-based methods.