Research On Kernel Learning Adaptive Modeling And Control For Industrial Batch Processes

The biochemical, fine chemical industries and many other batch-based production processes are very technology-intensive, and thus play an important role in the national economy. The product quality of batch process is usually very difficult to model and control. Consequently, this causes great quality fluctuations and restricts the development of the modern fine chemical industry. This dissertation focuses on the modeling and control of batch processes, mainly including the fermentation process and the rubber mixing process. This project will improve the technical state for batch processes, and enhance the application trend of information and calculating theory in the traditional chemical industry. Therefore, the topic of this thesis is to develop efficient, simple and suitable algorithms, using the recursive kernel learning theory primarily, for modeling and control of batch processes.The main contributions of this thesis are as follows:(1) A general and simple online modeling method, namely Selective Recursive Kernel Learning (SRKL), is proposed for nonlinear multi-input-multi-output (MIMO) processes. A two-stage recursive learning framework is first formulated for online identification. The samples can be introduced into and/or deleted from the model adaptively with a small computation load. Then, a general sparsification criterion for model complexity-controlled is developed to guarantee all the output channels of the MIMO model have good identification performance simultaneously. A novel pruning approach based on the fast leave-one-out cross validation criterion is explored to acquire controlled generalization ability of the identified model by deleting the useless information recursively. Consequently, the model can capture the process time-variant and dynamic characteristics by adjusting the model structure adaptively. The SRKL method is applied to online modeling of the penicillin and streptokinase processes. The obtained results show that the proposed SRKL method can online predict the biomass concentration and other key variables accurately. Furthermore, the model precision can be improved due to the process information accumulated batch-to-batch. The comparison study also shows that the SRKL method is superior to other common recursive kernel learning methods and traditional modeling methods, e.g., neural networks and recursive partial least squares. The benefits of identification accuracy, reliable performance and simple implementation in practice indicate that the SRKL method is very suitable for online identification of nonlinear systems.(2) A new online modeling method based on adaptive local kernel learning is proposed for nonlinear MIMO processes. Both distance measure and angle measure are adopted to evaluate the similarity between data, thus a more comprehensive and relevant set is constructed. In the present method, the model parameters are optimized online using the fast leave-one-out cross validation criterion. In addition, an online adaptive model selection strategy for modeling of batch processes is developed. Simulations on a fed-batch streptokinase fermentation process show that the active biomass and streptokinase concentrations can both be predicted simultaneously just from the second batch. The results also show that the proposed method is more accurate and adaptive, compared to traditional global and local kernel learning algorithms.(3) A novel control framework based on Sparse Kernel Learning One-step-ahead Predictive Control (SKL-OPC) is presented for general nonlinear processes. Two simple nonlinear controllers, namely SKL-PKR (Polynomial kernel & Root) and SKL-ATL (Adaptive Taylor Linearization) are designed, respectively.(3A) SKL-PKR is a direct controller suitable for the polynomial kernel. The manipulated input can be separated out from the control performance index due to the special structure of polynomial kernel. Consequently, the control law is obtained by solving the roots of an odd-degree polynomial equation. The proposed control strategy does not require the nonlinear optimization technique, which results in a small computation scale and makes it suitable for real-time control. A benchmark study is first presented to show that the proposed SKL-PKR controller is superior to traditional controllers and other related SKL model based control schemes. Then, application of the SKL-PKR controller to a highly nonlinear stirred tank reactor is explored in depth and the results compared to a well tuned proportional-integral-derivative (PID) controller indicate the advantages of this novel control strategy, in both deterministic and stochastic environments.(3B) SKL-ATL is an adaptive controller suitable for all kernel functions. The analytical control law is derived from Taylor linearization method. The convergence analysis of the SKL-ATL controller is presented based on the mean-value theorem, meanwhile a novel concept of adaptive modification index is obtained to improve its tracking ability and reject the unknown disturbance. The great advantage of SKL-ATL controller is its adaptability which overcomes the shortcoming of the other controllers, i.e., the tuning parameters are difficult to choose. Simulations on two benchmark problems and a highly nonlinear reactor indicate that the SKL-ATL controller exhibits fast and excellent tracking performance, meanwhile possesses satisfactory robustness and adaptability. The results also show that the proposed SKL-ATL controller is superior to the traditional PID controller and other related SKL model based control strategies.(4) The SKL-OPC framework is extended to the recursively updated form via the SRKL algorithm (Take the SKL-ATL controller for example). By updating the identified model, the SKL-ATL controller can trace the process characteristics and achieve better performance. Applications of the SKL-ATL controller to a nonlinear liquid-level system and a fermentation process show that it can exhibit more satisfactory performance, compared to the traditional PID controller and the corresponding control strategy without online updating.(5) Online measurement of Mooney viscosity is very difficult while critically important since Mooney viscosity can be likened to a composite measurement of the viscoelastic behavior of an elastomer and indirectly represents molecular weight and thus it has significant impact on end use properties of the polymer. A new modeling and control method based on adaptive RKL is proposed for online prediction and control of Mooney viscosity in rubber mixing processes. The model is established for each recipe and recursively updated to adapt the process fast changes. In the present method, a novel error evaluation index is proposed by combining the property of rubber mixing processes. Then, the model parameters are online adaptively selected, using the fast leave-one-out cross validation criterion, to overcome the shortcoming of user-defined parameters. An industrial system, named as Smart Mixing information integrated & control System (SMS), has been developed and successfully applied to several large-scale rubber and tire manufactories in China. The results of Mooney viscosity online prediction and the end discharge control show that the developed SMS is very efficient, and thus has real economic importance for industrial rubber mixing processes.(5A) The Mooney viscosity information can be obtained at the end of mixing process using such an adaptive model, without resorting to a long-term laboratory assay. Therefore, the intensity and pressure of lab assay work during mixing processes can be alleviated. Furthermore, the Mooney meters and related purchase cost can be decreased.(5B) Accomplishment of advanced discharge control of Mooney viscosity can achieve the consistent results of product, and thus improve the quality. The mixing duration can be lessened, consequently the energy is saved and the emission is reduced. Finally, the economic benefit of the rubber and tire manufactories can be promoted.(5C) Most of the production and process information of the rubber mixing process are integrated into SMS. Therefore, control and technique engineers can manage the process and moreover improve product quality with the assistance of SMS. The developed SMS will improve the technical state and enhance the application trend of automation and information theory in the rubber mixing process.Finally, the kernel learning based modeling and control methodology are addressed, including algorithms and industrial applications. The primarily theoretical background and philosophical inspiration are discussed. Some personal views & industry experiences are given for share. Some future research areas are highlighted.