Latent Variable Regression for Process and Quality Modeling

The supervised learning methods, partial least squares (PLS) and canonical correlation analysis (CCA), have been widely used in industrial processes to perform multivariate statistical modeling and monitoring based on process variables and quality variables. However, the latent variables extracted by PLS may contain irrelevant components, while CCA focuses only on the correlation but ignores the variance information. To overcome their drawbacks, a latent variable regression (LVR) modeling method with regularization is proposed to retain the prediction efficiency of CCA while exploiting the quality variance structure. LVR minimizes the prediction error between input and output scores, and retains consistent objectives in inner and outer modeling. Synthetic case studies and the Tennessee Eastman process are used to demonstrate the effectiveness of the proposed algorithm.