Dynamic Non-Gaussian and Nonlinear Industrial Process Monitoring Using Deep Analysis of Hybrid Characteristics

Data-driven monitoring using principal component analysis (PCA) and independent component analysis (ICA) has been studied extensively for modern industrial processes. However, due to the hybrid dynamic, nonlinear, and non-Gaussian characteristics of variables, complex industrial processes are difficult to model accurately. In view of the above issue, we present a novel method named deep analysis of hybrid characteristics (DAHCs) for industrial process monitoring. Considering that different monitoring methods can extract different characteristics, DAHC builds a process monitoring model on the basis of dynamic inner PCA (DiPCA), kernel PCA (KPCA), and ICA using a deep serial structure. First, process data are decomposed using DiPCA to analyze the dynamic characteristics. Then, the residual of DiPCA is further decomposed by multilayer KPCA-ICA for simultaneously analyzing nonlinear and non-Gaussian characteristics. Finally, the merged Bayesian inference comprehensive statistic (MBICS) for the dynamic, nonlinear, and non-Gaussian subspaces is used for integrated monitoring of hybrid characteristics. Typically, existing researches on hybrid characteristics indicate that single-layer feature extraction is insufficient for analyzing complex information in industrial processes. Current hierarchical feature extraction methods have yet to simultaneously account for dynamic, nonlinear, and non-Gaussian characteristics. DAHC adopts a deep serial structure to analyze the hybrid dynamic, nonlinear, and non-Gaussian characteristics of industrial data. Its feasibility and effectiveness are validated via the Tennessee Eastman (TE) chemical process and a practical mineral processing of lean iron ore. The results show that the TE and mineral processes have the highest average fault detection rates (FDRs) of 84.47% and 99.42% compared with other related methods.