Quality-based Process Monitoring with Parallel Regularized Canonical Correlation Analysis

摘要

Canonical correlation analysis (CCA) is a representative quality-relevant algorithm of multivariate statistical monitoring method. Different from PLS, CCA focus on extracting the correction between input and output data, which brings strong predictive ability. However, CCA ignores the variance of variables, which makes it more sensitive to noises. In addition, CCA suffers from collinearity problems that exists in process and quality data. In order to solve these problems, regularized parallel canonical correlation analysis is proposed. Firstly, regularization terms are inserted into constraint conditions of objective function. To remove the effect of noises, l₀minimization is applied on the correction matrix to eliminate zero or near-zero terms. The original data is decomposed into four subspaces: correlation subspace, input-relevant subspace, output-principal subspace and output-residual subspace. Besides, four corresponding monitoring statistics and control limits are then developed. Finally, numerical simulations are used to indicate the effectiveness of regularized PCCA.