【作者】 王鲜芳；

【导师】 潘丰；

【作者基本信息】 江南大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 随着生物工程技术的迅速发展,生化工业在国民经济中的地位已越来越重要。然而由于生化过程反应机理复杂,具有非线性、时变性、模型的不确定性等特点,以及缺乏可靠的传感器用于过程变量的在线检测等原因,其自动化水平与其它工业生产过程相比还远未成熟。因此,探求合适的智能建模方法和优化控制策略,并将其应用于生化过程已成为生化工程领域一个重要的研究方向,这对于促进生化工程技术的发展,降低原材料和动力的消耗,提高经济效益有着极其重要的意义。本论文的资助项目(863高技术重点发展项目子课题)“基于代谢工程和智能工程的新式、集约型发酵过程控制的工程化应用研究”就是针对生化工业的迫切需要而立项的,该项目研究的主要内容包括:1)研究以代谢模型为基础的发酵过程状态预测和推定、软测量、代谢流外部调控的普遍规律和关键技术;2)以代谢网络模型为基础的发酵过程在线控制技术的应用化研究;3)基于智能模型的发酵过程控制;4)研究基于智能模型的发酵过程数据采掘、多变量解析和在线故障诊断系统的关键技术,构建基于智能模型的发酵过程在线故障诊断/早期预警系统;5)过程控制示范装置和系统(报表、实时曲线、数据库等)。本课题针对1)、3)和5)进行研究,涉及的生化过程主要是谷氨酸补料分批发酵过程,所做的主要工作如下:通过对生化过程建模和优化控制的国内外研究现状以及存在问题的分析,在研究常用生化模型的基础上,利用标准支持向量回归(SVR,Support Vector Regression)算法对谷氨酸发酵过程实施软测量建模,通过实验验证该模型能够完成对谷氨酸浓度、菌体浓度和残糖浓度三个不能在线测量的状态变量的预估。从预测结果可以看出尽管模型达到了一定的预测精度,但发现该模型不能在线学习,而且当数据量增大时,还存在训练时间过长的问题。针对SVR软测量建模存在的问题,在充分研究最小二乘支持向量回归LSSVR (Least Square Support Vector Regression)算法的基础上,对谷氨酸发酵过程建立了基于LSSVR的软测量模型。通过实验发现,该模型尽管缩短了训练时间,但模型不能在线学习的问题依然没有解决。针对LSSVR软测量建模存在的问题,通过研究在线最小二乘支持向量回归OLSSVR(Online LSSVR)算法,提出了带遗忘因子的MIMO-OLSSVR(Multi-input Multi-output OLSSVR)算法,并在谷氨酸发酵过程软测量建模中进行应用。为提高算法的自动化程度,采用免疫遗传算法IGA(Immune Genetic Algorithm)使MIMO-OLSSVR的参数在建模过程中能够自动选择。将所建的模型预估谷氨酸发酵过程,结果表明该模型在保证较短训练时间的情况下,通过模型在线学习,实现了谷氨酸浓度、菌体浓度和残糖浓度三个变量的在线同时预测,提高了预测精度。但以上建模方法属于黑箱建模,不能反映实际生化过程的操控特点。为了能进一步探索生化过程的动力学特性,本文将混杂系统理论应用于谷氨酸补料分批发酵过程中,并建立了相应的混杂动力系统模型。在分段连续空间上,从理论角度证明了系统解的存在性、唯一性及解对初值和参数的连续依赖性。然后以产量最高为优化目标,利用混沌和单纯形法搜索合适补料操作变量。通过实验室的多次实验,证明该方法达到了提高产量的目的,为进一步研究生化过程控制提供一种新的理论思路。针对实际生化过程一般的数学模型或软测量模型常常随着时间的推移会发生模型失效问题,合理利用离线数据和在线数据,充分利用各种模型的优点,提出了多模型融合建模方法,实现了对谷氨酸发酵过程的动态建模。通过研究标准粒子群优化算法、量子粒子群优化算法,针对项目要求的多目标优化问题,提出改进的多目标量子粒子群优化(MQDPSO)策略,并用Deb测试函数进行测试。然后结合多模型融合建模技术,对谷氨酸补料分批发酵过程实施动态优化,通过实际应用验证了该方法的可行性。将改进的建模技术和控制策略融入集散控制系统中,能够进行实时状态预报和监控,实现了对谷氨酸补料分批发酵过程的建模和优化控制软件的模块化设计,便于在实际控制系统中的应用。更多还原

【Abstract】 The rapid development of biotechnology has highlighted biochemical industry in the national economy. However, its level of automation is still lower than other industrial processes because of the complex biochemical reaction mechanism, and the features of non-linearity, time variability, model uncertainty as well as the lack of reliable sensors to be used in on-line inspection of process variables. For these reasons, the studies of suitable intelligent modeling, optimization control strategy and the application to biological and chemical process have become a new trend in biochemical engineering, which may be of great profundity promoting the development of chemical and biological technology and reducing the consumption of raw materials and energy.The funding program for this paper (863 program focusing on the development of high-tech sub-topic) " The Applied Engineering Research of Modern, Intensive Fermentation Process Control Based on Metabolic Engineering and Intelligent Works " is intended to solve the urgent problems for biochemical industry, which includes: 1) the study of the state prediction and presumption, soft sensor, the universal law of external regulation of metabolic flow and key technologies of the fermentation process based on the metabolic model; 2) the study of application research of online controlling technology of fermentation process based on the metabolic network model; 3) the study of controlling for fermentation process based on the intelligent model; 4) the study of critical technologies for data mining, multivariable analysis and online fault diagnosis in the process of fermentation based on intelligent model to construct an on-line fault diagnosis or early warning systems;5) the study of methods to construct demonstration devices and systems (reporting, real-time curve, databases, etc.) for the process controlling. Of the 5 aspects of the study, my research focuses on 1), 3) and 5), of which the biochemical processes involved is mainly about glutamic acid fed-batch fermentation process. The main done work goes as the follows:A soft model of glutamic acid fermentation process is constructed by using the standard support vector regression (SVR) on the basis of common biochemical models and analyzing current domestic and international researches of biochemical process modeling and optimization controlling. The experiments verify that this model can effectively measure 3 state variables, namely, glutamic acid concentration, biomass concentration and residual sugar concentration, which can not be done on-line. However, although the forecast model has a certain degree of accuracy, but somen problems are still existed, such as the model can not be training online, and the training is more time-consuming with the increasing sampling number.To solve the problem of the soft-sensor model using the SVR, another soft modeling based on least square support vector regression(LSSVR) algorithm for glutamic acid fermentation process is constructed. The experiments indicate that the model despite shortened training time the prediction accuracy declines, the problem that the model can not be training online still remain.For the problems with the soft-sensor modeling using LSSVR, an improved multi-input and output online LSSVR (MIMO-OLSSVR) with a forgetting factor is designed and applied. To improve the automation degree of the algorithm, an immune genetic algorithm (IGA) is adopted to achieve automatic selection of MIMO-OLSSVR parameters. Using the MIMO-OLSSVR model to estimate glutamic acid fermentation process, the experiments show that the model can simultaneously predict the three state variables and improve the accuracy while reducing time consumption by training online. However, this modeling is a black box in nature and does not reveal the operational characteristics of the actual biochemical process.To further explore the dynamics of biochemical process, the theory of hybrid systems is adopted for glutamic acid fed-batch fermentation process, and a hybrid kinetics system model is accordingly constructed. The research has proved the existence of systematic solution, uniqueness and the continuous solution dependence on the initial values and parameters in the space of continuous segmentation. Taking the highest production as the optimization target, and searching for the suitable feeding operation variables by utilizing the chaos and simplex method, the experiments prove that the model can achieve the purpose, which may provide a new theoretical for the research of biochemical process control.This study proposes a dynamic multi-model fusion modeling method for glutamic acid fermentation process by using online data and offline data on the basis of various models to solve problems from weakness of the mathematical model or the soft-sensing model, which may often fail in actual biological changes with the time advance. For the multi-objective optimization as the program requires, an improved multi-objective quantum delta-potential-well-based particle swarm optimization (MQDPSO) strategy is presented on the basis of standard particle swarm optimization algorithm and is tested by the Deb test functions. Then combined with the techniques of multi-model fusion modeling, dynamic optimizing for glutamic acid fed-batch fermentation process is achieved and the feasibility of the optimization method is proved.The improved modeling techniques and the control strategies integrated into the distributed control system can achieve the purposes of real-time status of forecasting, monitoring and modularizing modeling and optimizing the glutamic acid fed-batch fermentation process, which can bring much convenience to the application in real control systems.更多还原