Chemical Process Monitoring Methods Based On Global-local And Discriminant Feature Extraction

Feature extraction is a key step that affects monitoring performance in data-driven process monitoring methods,and it is a hot research topic in process monitoring.As one of the most widely used feature extraction algorithms,the multivariate statistical algorithm,represented by the principal component analysis and its extensions,focuses on the global features represented by different order statistics of the data,ignoring the local features represented by the neighborhood information of the data,which leads to certain important features being covered by noise and the fault detection performance being compromised.Therefore,global-local feature extraction algorithms are proposed to integrate multivariate statistical algorithms with manifold learning algorithms to incorporate local features of data,so as to achieve comprehensive extraction of data features.Furthermore,how to apply the above feature extraction algorithms to processes with different process characteristics is the focus of current research,but the deviations of data feature description and measurement have not been specifically discussed in their implementation.In addition,compared with the traditional process monitoring methods oriented on normal condition data,the discriminative information in historical fault data has not received sufficient attention.With these concerns,the following research is proposed in this thesis.In order to solve the non-uniformity of the data distribution,a process monitoring method based on improved global-local preserving projections is proposed.By constructing adaptive neighborhood and introducing geodesic distance,the description and measurement of data features are improved to enhance the fault detection performance.Furthermore,considering the influence of data curvature on its determined neighborhood relations,a novel global-local feature preserving projections algorithm is proposed in this dissertation.Curvature weight is introduced into the linear tangent space alignment algorithm to modify the estimation of adaptive neighborhood tangent space,by which the comprehensive extraction of local and global features is achieved.By comparing monitoring results of the above monitoring methods with those similar monitoring methods,the identification of process deviation can be significantly improved by proposed data feature extraction.To further consider the discriminative features contained in historical fault data,a process monitoring method based on global-local marginal discriminant preserving projections is proposed.Inherent feature extraction based on globallocal features is integrated with discriminative feature extraction based on multiple marginal Fisher analysis through Fisher’s criterion in this method.Specifically,historical fault data is used to enhance the extraction of discriminative features between normal conditions data and abnormal data,so as to obtain a feature space that could best distinguish normal conditions and fault,and further improve the historical fault detection performance.The proposed method is validated by comparing the monitoring results with conventional methods.