Kernel Canonical Variate Analysis for Nonlinear Dynamic Process Monitoring

Abstract

Effective monitoring of industrial processes provides many benefits. However, for dynamic processes with strong nonlinearity many existing techniques still cannot give satisfactory monitoring performance. This is evidenced by the well known Tennessee Eastman (TE) benchmark process, where some faults, e.g. Faults 3 and 9, have not been comfortably detected by almost all data-driven approaches published in the literature. This is because most data driven approaches, such as the principal component analysis (PCA) are linear. In recent years, powerful nonlinear analysis tools using kernel principles have been proposed. However, these tools have not been successfully applied to dynamic systems due to enormous dimensionality and complexity issues. This paper proposes nonlinear dynamic process monitoring based on kernel canonical variate analysis (KCVA). The proposed technique performs the traditional canonical variate analysis with KDE (CVA-KDE) in the kernel space generated from kernel PCA. The kernel PCA accounts for the nonlinearity in the process data while the CVA captures the process dynamics. The approach was tested on the TE benchmark problem for fault detection. The results obtained showed that KCVA detected faults at a higher rate and much earlier than CVA especially in the more difficult faults such as Faults 3 and 9 in the TE process which cause very little variation in the measured variables.