Please use this identifier to cite or link to this item:
|Title:||DATA-DRIVEN MODELING FOR QUALITY CONTROL IN CHEMICAL PROCESSES|
|Abstract:||This thesis considers the problem in data-driven modeling for quality control of complex chemical processes characterized by nonlinearity and time-varying dynamics over a wide range of operating conditions. The critical steps in data-driven modeling for quality control, such as data pretreatment, online inferential model maintenance and quality control mechanism designs, are investigated separately in previous studies, which calls for holistic research on those topics. In the initial phase of this research, the missing measurement problem in data-driven modeling is addressed by two developed probabilistic partial least squares (PPLS) methods. In the expectation-maximization (EM) based PPLS (EM-PPLS), the convergence properties for model parameters and estimated values of missing measurements are proved in explicit forms, enabling effective and robust application. In parallel, full Bayesian treatment is employed in the Bayesian inference based PPLS (BI-PPLS), to substantially reduce the computational complexity in latent variable selection in the PPLS model. In both proposed PPLS methods, data-driven model generation and missing measurement estimation can be carried out concurrently along with the uncertainty handling in process measurements through a probabilistic strategy. Next, the degradation problem in the online implementation of the data-driven inferential models is addressed. The discrepancy between the actual model parameters in the data-driven model and filtered model parameters by Kalman filter is calculated and utilized to derive a model mismatch index, decomposition of which can further lead to identification of leading abnormal model parameters in the data-driven model. With the developed model mismatch index, an online model update strategy is developed to maintain the data-driven inferential model online. Finally, a predictive controller based on multiple data-driven dynamic and inferential models is designed to control the final product quality in the batch processes. The entire batch is divided into several modeling phases based on the availability of intermittent quality measurements, in which local dynamic models are identified using weighted autoregressive exogenous (ARX) models. Meanwhile, multiway PLS (MPLS) models are also constructed to predict the quality at batch termination given the intermittent quality measurements at the initiation of each modeling phase. The data-driven models are integrated within model predictive control (MPC) framework to control the final product quality in Penicillin fermentation process with the consideration of causality relationships between the future inputs and final product quality.|
|Appears in Collections:||Open Access Dissertations and Theses|
Files in This Item:
|thesis.pdf||Thesis||1.24 MB||Adobe PDF||View/Open|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.