Concurrent Quality-Relevant Canonical Correlation Analysis for Nonlinear Continuous Process Decomposition and Monitoring

The main focus of this work is to propose a data-driven residual generation based performance monitoring method for an industrial nonlinear p-xylene (PX) oxidation reaction process. In the proposed method, canonical correlation analysis is used as the basic method for residual generation and regression, because of its optimal detection capability in the sense of a given confidence level. Then, the process data are partitioned into several subspaces according to their properties with respect to the relevance to the final quality of the products. In these subspaces, the process can be monitored in a concurrent manner. Meanwhile, multikernel learning (MKL) is introduced to expand the nonlinear process data to a relative linear space in a reasonable way. The main contributions of this work are as follows: (1) The entire process is divided into four parallel subspaces and monitored concurrently, which means the proposed method can be applied distributively. (2) Considering that the PX oxidation reaction process is a highly nonlinear one, MKL is introduced to extend the concurrent canonical correlation analysis from linear cases to nonlinear ones. The performances of the proposed method are validated in an industrial PX oxidation reaction process.