Some Researches On Beneficial Information Concentration And Feature Selection-Based Multivariate Statistical Chemical Process Monitoring

In modern process industry, process monitoring and fault diagnosis are gaining significant attentions because of the increasing demand in plant safety and product quality. With the development of data gathering equipment and computing technology, process data information become very abundant, leading to the rapid development of data-driven chemical process monitoring and fault diagnosis. Especially, as the multivariate statistical analysis technique progresses, multivariate statistical process monitoring (MSPM) and fault diagnosis becomes a very hot topic in process monitoring field and has gained significant attentions in both academic and industrial circles.Although many research results and successful applications of MSPM have been reported, MSPM is still under the development, and many issues need to be further discussed and addressed. For example, how to extract more beneficial fault information for monitoring, how to explore more process local behaviors, and how to deal with the nonlinear, multimode process characters. Based on the existing results, this dissertation makes some further researches and improvements on the existing works and proposes several efficient monitoring methods from the aspect of beneficial information conceration and feature selection which are summarized as follows.(1) Considering the subjective PC selection and the beneficial information dispering problem in principal component analysis (PCA)-based process monitoring, the dissertation analyzes the PCA monitoring performance further and proposes PC selection and dominant subspace re-construction strategies for monitoring performance improvements. Because the most beneficial information has been concentrated into the same subspace, both fault detection and diagnosis performances are improved. Given the fundamentality of PCA, the proposed methods can be extended extensively to solve various practical monitoring problems.(2) Considering that the modern industries are usually characterized by large number of variables, complex variable correlations and multiple operation units, the dissertation proposes a totally data-driven multiblock monitoring method. Dividing the variables into different subspaces can not only reduce the system complexity but also explores more process local behaviors, leading to improved monitoring performance. The proposed mothod can be extended to nonlinear or multimode form to solve more monitoring problems.(3) Considering the useful information dispersing and being submerged problem in independent component analysis (ICA)-based process monitoring, the dissertation proposes a components pre-selection method and a double-weighted monitoring scheme respectively for performance improvements. The pre-selection ICA-based monitoring makes the best use of fault information and concentrates the most beneficial information into the same subspace; the double weighted ICA monitoring method evaluates the importance of each IC online and highlight the ICs representing the most variational information in the monitoring. The proposed method can be extended to nonlinear or multimode form to be more suitable for practical application.(4) Considering that the multiple variable distributions in plant-wide processes, the dissertation proposed a distribution similarity-based multiblock monitoring method which divides the variables with similar distribution into the same block. In the sub-blocks, the variables are more likely to be linear correlated and can be regarded as linear combinations of some latent variables, making ICA more suilable for monitoring in each sub-block. The proposed method is domenstrated to be effective in both numerical process and the Tennessee Eastman benchmark process.Finally, conclusions and future research studies of the multivariate statistical process monitoring areas are discussed.