Nonlinear chemical process monitoring using decentralized kernel principal component analysis and Bayesian inference

Traditional kernel principal component analysis (KPCA) based process monitoring method generally depends on a global monitoring model, however, because of the complex relationships among numerous variables in modern industrial processes, KPCA monitoring technique may not function well. Inspired by the recognition of this shortcoming, a novel totally data-driven multi-block statistical process monitoring method (MBSPM) based on evaluating nonlinear relations between variables is proposed. In MBSPM, a new nonlinearity correlation coefficient is firstly employed to divide process variables into several individual subspaces and each subspace shows strong nonlinear characteristic. Then, to cope with nonlinear correlations, KPCA is adopted to build process monitoring models respectively, thus, the constructed monitoring models in different subspaces can reflect the local behavior of certain faults. The results in all subspaces are combined together to create a final statistic by Bayesian inference. The feasibility and validity of the proposed method is finally demonstrated through comparison studies on the Tennessee Eastman process.