(446c) Non-Gaussian Dynamic Process Monitoring Using A New Dissimilarity Method Based On Independent Component Analysis and Multidimensional Mutual Information

Traditional multivariate statistical processes monitoring (MSPM) techniques like principal component analysis (PCA) and partial least squares (PLS) are not well-suited in monitoring non-Gaussian processes because the derivation of monitoring indices requires the approximate multivariate Gaussian distribution of the process data. To address the challenges of traditional MSPM methods, a novel pattern analysis driven dissimilarity approach is developed by integrating multidimensional mutual information (MMI) with independent component analysis (ICA) in order to quantitatively evaluate the statistical dependency between the independent component subspaces of the normal benchmark and monitored data sets.

In this study, ICA is integrated with MMI to measure the statistical independency between two independent component subspaces, which represents the subspace dissimilarity. The presented approach uses entropy based mutual information to assess the dissimilarity between the multidimensional independent component subspaces of the benchmark and monitored sets. The new MMI based ICA dissimilarity index is derived from the higher-order statistics so that the non-Gaussian process features can be extracted efficiently. The proposed approach does not require any assumptions regarding the relationships among measurement variables and can deal with the process data following an arbitrary probability density distribution. In contrast to the angle based dissimilarity factors, the new multidimensional mutual information based dissimilarity takes into account the higher-order statistics and thus can capture the non-Gaussian features of the process data well. Furthermore, a rolling window that moves over the monitored segment incrementally ensures that the time-varying process dynamics are accounted for in the MMI based dissimilarity approach. The higher-order statistics underlying the dissimilarity index can help extract the non-Gaussian features and quantify the statistical dependency between IC subspaces on a moving-window basis. Thus the non-Gaussian dynamic processes can be effectively monitored through the continuous comparison between the normal benchmark and monitored data sets.

The multidimensional mutual information based ICA dissimilarity method is applied to the Tennessee Eastman Chemical process. The comparison of monitoring results demonstrates that the new mutual information based dissimilarity index is superior to regular PCA, angle based PCA dissimilarity, conventional ICA and angle based ICA dissimilarity methods with the most reliable fault detection capability.