Fault detection of process correlation structure using canonical variate analysis-based correlation features

This paper proposes a canonical variate analysis (CVA) approach based on feature representation of canonical correlation for the monitoring of faults associated with changes in process correlations, which involves two new metrics, Rs and Rr, corresponding to the state and residual spaces. The utilization of the canonical correlation feature can improve the monitoring proficiency by providing more application-dependent representations compared with the original data, as well as a decreased degree of redundancy in the feature space. A physical interpretation is provided for the canonical correlation-based method. The effectiveness of the proposed approach for the monitoring of process correlation changes is demonstrated for both abrupt (step change) and incipient (slow drift) types of faults in simulation studies of a network system. In the simulation results, the canonical correlation-based method has superior performance over both the causal dependency-based method and the traditional variable-based method.