(714e) Independent Component Analysis Mixture Model Based Dissimilarity Method for Performance Monitoring of Non-Gaussian Multimode Dynamic Processes

Data-driven multivariate statistical methods have been developed to extract useful information from a large amount of process data and detect various types of process faults with the high-dimensional and correlated process data. In order to monitor multimode dynamic processes that have non-Gaussianity within each operating mode, an independent component analysis (ICA) mixture model based non-Gaussian dissimilarity method is developed in this study.

ICA mixture model (ICAMM) is integrated with mutual information based non-Gaussian dissimilarity index for performance monitoring of multimode dynamic processes with non-Gaussian features in each operating mode. The normal benchmark data set is assumed to be from different operating modes with non-Gaussian process features in each mode, which can be characterized by a localized ICA model. Thus an ICA mixture model is developed with a number of non-Gaussian components that correspond to various operating modes in the normal benchmark set so that the non-Gaussian structure is retained in each component. Then, a sliding window strategy is carried out to obtain a series of subsets of monitored data with the same length as the benchmark set for taking into account the process dynamics. Further, each sample in the subset of monitored data is classified into a local ICA component through the maximized Bayesian posterior probability. Then the statistical independency between the IC subspaces of the benchmark set and the monitored subset corresponding to the local operating mode are estimated as the dissimilarity factor to evaluate the likelihood of the monitored operation to be abnormal. The higher-order statistics of entropy and mutual information are taken into account and thus non-Gaussian process features in shifting process operation conditions are captured from the proposed ICA mixture model based dissimilarity index by comparing the localized IC subspaces between the benchmark and the monitored data sets.

The proposed ICA mixture model based dissimilarity method is applied to monitor the performance of the Tennessee Eastman Chemical process with multiple operating modes and the fault detection results demonstrate the superiority of the proposed method over the conventional eigenvalue decomposition based and geometric angle based principal component analysis (PCA) mixture dissimilarity methods. It is shown that the new ICA mixture model dissimilarity method has strong capability of detecting process faults while mitigating false alarms for monitoring the performance of multimode non-Gaussian processes.