Complex process quality prediction using modified kernel partial least squares

In this paper, kernel partial least squares (KPLS) method is modified based on orthogonal independent component analysis (O-ICA). Then it is applied to quality prediction of industrial processes. In ICA, the extracted components are assumed to be mutually statistically independent instead of uncorrelated. Independence is much stronger than uncorrelativity. Those extracted ICs may thus provide more informative statistical explanations and better reflect the inner properties of measurement data. However, disturbing variation can be extracted since ICA uses entropy theory to extracts high-order statistics. Hence, first, O-ICA is proposed for signal correction of non-Gaussian processes. Then KPLS is modified for quality prediction of non-Gaussian processes based on O-ICA, which is called O-ICA-KPLS. Advantages of the proposed O-ICA-KPLS are: (1) has the ability to give high-order representations for non-Gaussian data compared to original KPLS, and (2) provides more accurate statistical analysis and on-line monitoring because independent signals are corrected. The proposed methods are applied to the quality prediction in fermentation process and Tennessee Eastman process. Applications indicate that the proposed approach effectively captures the relations in the process variables and use of O-ICA-KPLS instead of original KPLS improves the predictive ability. (C) 2009 Elsevier Ltd. All rights reserved.