Multivariate statistical process monitoring and its integration with HAZOP analysis for abnormal event management

Fangping Mu, Purdue University

Abstract

Abnormal event management (AEM) is an important problem in industrial chemical process operations. Principal component analysis (PCA) is widely used for process monitoring, which is the first step of AEM. The PCA method is sensitive to outliers in the data used to build the model. Robust PCA is first proposed using robust estimator as a diagnostic tool. Minimum covariance determinant (MCD) estimator is used as a diagnostic tool for robust PCA. A hybrid genetic algorithm is proposed for MCD estimation. After outliers are removed, PCA is built based on the remaining ‘good’ part of the data. To provide a tradeoff between false alarm and quick detection, a novel statistical testing algorithm is integrated with PCA to improve the fault detection and identification performance. Scores space and residuals space generated by PCA is decomposed into several subsets so chosen that in each subset the detection problem can be solved with an efficient recursive change detection algorithm based on χ2-generalized likelihood ratio (GLR) test. PCA is a global linear model. For a highly nonlinear chemical process, nonlinear PCA based on local linear approximation is proposed for process monitoring. Integration of Isomap, an efficient algorithm to generate the intrinsic dimensionality of the process data, with mixture of probabilistic principal component analyzers, is proposed to generate nonlinear PCA model, which can then be used for process monitoring. HAZOP analysis is a systematic proactive identification, evaluation and mitigation of process hazards during the design stage of a process. PHASuite, a knowledge based system for automated HAZOP analysis, is overviewed. A framework to integrate PCA and PHASuite for online abnormal event management is then proposed. The framework includes three major parts: process monitoring, automated HAZOP analysis module and a coordinator. Multiblock PCA is used for process monitoring of continuous process, while multiway PCA is used for batch process. When an abnormal event is detected, quantitative to qualitative measurements transformation based on contribution plots is then proposed. Based on the identified qualitative states, HAZOP analysis is performed using PHASuite based on digraphs to locate the original deviation, potential causes, consequences and recommendations to mitigate the consequences.

Degree

Ph.D.

Advisors

Venkatasubramanian, Purdue University.

Subject Area

Chemical engineering

COinS