【作者】 冯絮影；

【导师】 王建林；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 微生物发酵是生物工程和现代生物技术及其产业化的基础。随着生物工程技术的进步和发酵工业生产规模的不断扩大,迫切需要提高发酵过程控制系统的控制性能和鲁棒性。发酵过程是典型的非线性、非平稳、高维数、慢时变的复杂系统,且缺乏足够的先验知识,难以构建准确描述发酵过程特性的数学模型。随着对发酵过程控制性能要求的不断提高,研究能准确表征发酵过程特性的混合建模方法及有效的自适应控制策略具有重要的理论意义和应用价值。本文在分析现有发酵过程建模和控制方法研究现状基础上,以系统的未知动态不确定性为主要研究对象,针对模型失配和未知动态等不确定性影响发酵过程控制品质的难题,研究了发酵过程的混合模型结构、在线快速建模、状态抗差估计、鲁棒预测控制等问题,并进行了青霉素发酵过程预测控制的仿真实验研究。发酵过程机理十分复杂,已有的建模方法对过程的描述并不全面。本文提出了一种带未知动态的非线性系统混合状态空间模型,未知动态的引入使得该模型不但具备一般非线性系统的优点，还能表征更广泛领域的非线性过程特性。在此基础上,设计了带动态补偿的输出反馈非线性模型预测控制(Nonlinear Model Predictive Control, NMPC)的整体实现框架,并分析了未知动态对系统动态特性的影响,为后续研究打下了坚实的理论基础。基于核方法的在线建模研究方面,为提高支持向量回归(Support Vector Regression, SVR)建模的实时性和运算速度,本文首次从在特征空间复制训练样本的角度证明了SVR和支持向量分类(Support Vector Classification, SVC)的等价性问题,部分消除了SVC和SVR训练算法的差异,将简洁快速的几何训练算法推广用于SVR训练,有效提高了SVR的实时性和运算速度。基于核独立元分析(Kernel Independent Component Analysis, KICA)和递推最小二乘支持向量机(Recursive Least Square Support Vector Machine, RLS-SVM)的未知动态回归估计满足了在线建模和预测估计的实时性要求,且具有较低的计算复杂度。基于滤波器的自适应控制器设计方面,基于过程混合模型,研究了基于环路传递复现(Loop Transfer Recovery, LTR)非线性二次调节(Nonlinear Quadratic Regulation with LTR, NQG/LTR)的、带一步动态补偿的NMPC (NMPC with Dynamic Compensation, NMPC/DC)设计方法,考虑到高斯和非高斯扰动的情况,本文在不敏粒子滤波(Unscented Particle Filter, UPF)和不敏变换抗差Kalman滤波(Unscented Transformation based Robust Kalman Filter, UT-RKF)状态估计的基础上,构建了完整的基于NQR/LTR的发酵过程NMPC/DC控制系统,并给出了控制系统的详细执行步骤。以青霉素发酵过程为实验对象,仿真实验表明：用基于不同训练算法的SVR对青霉素发酵过程大规模数据进行回归训练,在相同的实验条件和相当的回归精度情况下,基于几何训练算法的SVR具有更快的运算速度和更好的收敛性,有效提高了SVR的数据实时处理能力；在初始值和噪声方差存在偏差的情况下,比较EKF、RKF、UT-RKF和RPF的状态估计性能可知,UT-RKF和RPF算法具有更好的估计精度和数值稳定性,高斯扰动情况下,UT-RKF可在计算复杂度较低的情况下得到较好的状态估计结果,便于在线计算；同开环控制、不带未知动态补偿的控制和只考虑未知动态补偿的控制相比,提出的基于模型在线更新的发酵过程NMPC/DC控制能很好的跟踪最优轨迹,具有更好的控制性能和鲁棒性。本文提出的基于模型在线更新的NMPC/DC系统具有坚实的理论基础,大幅度提高了控制系统鲁棒性和控制性能,为复杂非线性控制系统性能的提高提供了有效的途径,具有重要的指导意义,在化工、机电等复杂非线性领域有着广泛的应用前景。更多还原

【Abstract】 Fermentation is one of the most important technical elements of biological engineering and modern biotechnology. Because of the progress on biotechnology and increasing scale of industrial fermentation production, it became an urgent need to improve the fermentation process control performance and system robustness. Fermentation process is typically non-linear, non-stationary, high-dimension, time-varying, and lack of adequate prior knowledge. It is difficult to establish an accurate mathematical model to describe the process characteristics. As advancement of the fermentation process control performance requirements, the studies on hybrid modeling methods and efficient adaptive control strategies have important theoretical significance and potential application value.In this paper, the existing fermentation process modeling and process control methods are analyzed in detail. Aim at the problems in fermentation process control, the nonlinear systems with uncertainties, which include model mismatching and unknown dynamics, were mainly studied. The research issues include hybrid model structure, on-line fast modeling, robust state estimation, robust predictive control, etc. A simulation study of Penicillin fermentation was made to confirm the proposed methods.As the fermentation process is very complex process, it is difficult to obtain good results based on traditional modeling method. Therefore, the paper presents a hybrid state space model to describe a nonlinear system with uncertainties. By introducing local dynamics, makes the model not only has the advantages of general nonlinear systems, but also characterized a wider field of nonlinear process. This paper discusses the stability of the nonlinear system with unknown dynamics, and the corresponding compensation methods with an output feedback nonlinear model predictive control (NMPC). The overall implementation framework laid a solid theoretical basis.In the kernel-based online modeling studies, in order to enhance the support vector regression (SVR) modeling speed, this paper presented the geometric interpretation of support vector machine (SVM). It was proved that SVR and support vector classification (SVC) are the equivalence problem by duplicating the training samples in the feature space. Some of the geometric training algorithms were introduced to speed up SVR training, which are simpler and faster than other QP-based algorithms but only suit for SVC earlier. In addition, kernel independent component analysis (KICA) was introduced to decomposition unknown dynamics of the independent impact, and a novel in-place moving-window recursive least square SVM (RLS-SVM) algorithm with lower computational complexity was used to model and forecast the unknown dynamics.In the studies of filter-based adaptive controller design, the remodeling strategies based on hybrid model were discussed firstly. The controller design methods based on nonlinear quadratic regulator with loop transfer recovery (NQR/LTR) were studied. Further, based on robust particle filter (RPF) and unscented transformation based robust Kalman filter (UT-RKF) algorithm, two dynamic compensator inbuilt NQR/LTR algorithms were proposed, and gave the complete adaptive output feedback predictive control algorithm system of fermentation process.By the penicillin fermentation process control simulation experiment, the results show that, for SVR training on large-scale data in the same experimental conditions, the regression based on geometric algorithm was quite precise with faster computing speed and better convergence, and effectively improve the SVR for real-time processing. The comparative studies of EKF, RKF, UT-RKF and RPF for state estimation under different initial value, noise variance cases show that, UT-RKF algorithm has higher estimation accuracy and better numerical stability when the noise is Gaussian, and UT-RKF has lower computational complexity. Comparing with the open-loop control and NMPC without dynamic compensation, the proposed dynamic compensator inbuilt NMPC could track the optimal trajectory very well, and have a better control performance and robustness.This paper presents a dynamic compensator inbuilt NMPC system for fermentation process, it has good robustness and control performance, provides a significance method to improve the complex non-linear process control system performance, and has a broad application prospects.更多还原