【作者】 崔蕾；

【导师】 袁景淇；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 维生素C(又名L-抗坏血酸,L-ascorbic acid)是人体必需的一种维生素,我国维生素C生产采用的是具有自主知识产权的二步发酵法生成其重要前体—2-酮基-L-古龙酸(2-keto-L-Gulonic acid, 2-KGA),2-KGA经过烯醇化酸化生成维生素C。随着生物发酵技术的发展以及生产规模的不断扩大,工厂对发酵过程的自动化水平、安全性和可靠性要求越来越高。然而,2-KGA发酵过程不同于一般的化工过程,它涉及两种菌体的繁殖、代谢过程,两种菌体之间的生长相互影响,机理复杂,难以用准确的数学机理模型来描述。随着计算机集散控制系统的应用与发展,大量的过程数据被及时采集和存储,如何利用这些数据包含的过程状态信息提高发酵过程性能是当前过程优化控制领域的研究热点问题,也是本文的研究内容。本文基于2-KGA发酵过程丰富的在线测量和离线分析数据,利用人工智能技术(支持向量机、智能数据库技术、模糊逻辑等)和统计分析方法,研究建立了基于数据驱动的优化2-KGA发酵过程的关键技术,首先对发酵过程关键状态变量进行超前预报,其次建立2-KGA发酵过程的评价指标—效益函数并对其进行超前预报,在预报的基础上在线评估发酵批次的创利潜力,从而对多罐并行发酵系统实施优化调度,最后在上述研究基础上开发了一套相应的软件在发酵生产车间在线运行,形成2-KGA发酵过程运行优化体系。主要研究成果包括:1.建立2-KGA发酵过程关键状态变量的超前预报模型2-KGA发酵过程关键状态变量(如产量、2-KGA浓度)难以在线测量,但反映了过程所处的状态,是优化发酵过程的重要依据,因此状态变量的超前预报将对发酵过程的优化有指导意义,本文建立2-KGA发酵过程产量超前预报模型,将AdaBoost回归模型与基于支持向量机滚动-学习预报技术相结合,依次训练一组基于支持向量机的滚动学习-预报器(弱预报器),每个弱预报器采用不同的参数,利用AdaBoost算法将弱SVM预报器提升为强SVM预报器,减少了调试参数的工作量,利用2-KGA工业发酵过程数据对本文建立的预报模型进行验证表明其具有较好的泛化性能和抗噪性能。在产量预报的基础上,利用统计分析对发酵液体积进行预估进而得出2-KGA浓度的预报值。另外,生产过程染菌的难以避免造成异常批次的产生,本文建立基于产量预报的异常批次早期发现方法,利用工业发酵数据进行验证表明该方法是可行的,能及时给出预警信息。2.建立2-KGA发酵过程效益函数的在线计算与预报模型对2-KGA发酵车间实施优化调度首先需要建立描述2-KGA发酵过程性能的评价指标,本文建立效益函数这一指标反映过程的经济行为,效益函数定义为单位时间内一个批次所创造的毛利润,通过物料衡算、能量衡算在线计算出发酵批次的效益函数,并采用基于SVM的滚动-学习预报技术对其进行超前预报,为发酵批次的在线分类、优化调度作准备。首先给出用于效益函数预报的初始训练库的建立,然后提出了历史数据库的两种在线更新方案,即分别基于分类、K-NN算法的更新策略,进而利用2-KGA工业发酵过程数据对两种方案的预报性能进行了评估,确定出预报效果较好的更新数据库方案。3.对2-KGA发酵过程实施优化调度本文对2-KGA工业发酵过程建立的优化调度策略主要是提供一种补料优化方法,改变传统的、各批次补料量固定(L-山梨糖资源平均分配)的生产方式,优化山梨糖在各个运行批次中的分配方案。首先根据效益函数的实际值和预报值对发酵批次在线分类,实时评估发酵过程的运行状态,获得发酵过程的创利潜力,利用状态反馈信息来确定或调节该批次的补糖量。创利潜力高的优势批次通过继续流加碳源(L-山梨糖)适当延长操作周期,挖取该批次创利潜力,而对创利能力差的劣势批次,适度减少碳源分配量,避免发酵车间的效益损失,从而使多反应器并列运行车间的经济效益最大化。此外,文中还讨论了实施优化调度方案时基于模糊逻辑技术的发酵周期的预估问题。通过对背景厂发酵车间的历史批次实施拟在线优化调度,该调度方案将给车间带来6.86%的收益增量。4.开发了2-KGA发酵过程在线预报-优化调度软件在上述研究的理论成果基础上,开发了一套2-KGA发酵过程在线预报-优化调度软件,该软件实时采集发酵过程各种数据(在线测量和离线分析数据),采用支持向量机、AdaBoost算法、统计分析、智能数据库、模糊逻辑等数据驱动技术,实现2-KGA发酵过程产量、产物浓度及效益函数的在线预报、发酵批次创利潜力的在线评估,以及发酵车间优化调度、发酵过程阶段辨识和异常批次的早期发现等优化控制的关键技术应用于工业发酵生产中,提高了2-KGA工业发酵过程的自动化水平。该软件已在背景厂发酵车间安装调试成功并在线运行。更多还原

【Abstract】 L-ascorbic acid (L-AA) also known as vitamin C is an essential nutrient for human being. In the 1970s, an innovative bioconversion process was developed for the production of 2-keto-L-gulonic acid (2-KGA), a key precursor for L-ascorbic acid synthesis. With this process, 2-KGA is produced via the two-stage fermentation, and then converted to L-AA by catalytic reactions. Since then, fermentation became the dominant process in the L-AA production. In the last decades, the equipment and production scale in fermentation industries are rapidly expanded with the development of biotechnology, then the higher automation, stability and reliability of the process are required. However, it is difficult to obtain the kinetic modeling for 2-KGA fermentation processes because the complex interactions of the two involved microorganisms are not well known yet. On the other hand, a large amount of process data are collected from industrial 2-KGA cultivation and the data contain abundant information of processes. Therefore, data-driven improvement of the process performance is focused on in this dissertation.Based on the on-line measurements and off-line assay data, key technologies are studied in this dissertation for optimizing industrial 2-KGA fermentation processes, such as on-line prediction of state variables, optimal scheduling for multi-bioreactor workshop, process monitoring and fault detection, where artificial intelligence technologies (e.g., support vector machines, intelligent database and fuzzy logic) and statistical analysis methods are applied. The main contents are as follow:1. Prediction of state variables in industrial 2-KGA fermentation processes The state variables such as 2-KGA formation and 2-KGA concentration could provide important information for the optimization of fermentation processes. Data-driven prediction of the state variables is presented in this paper. The 2-KGA formation is on-line predicted by integrating SVM-based rolling learning-prediction technology with the AdaBoost algorithm. The AdaBoost algorithm is used to adaptively boost the performance of SVM predictors, which is demonstrated to be beneficial to improve the prediction accuracy and the robustness. The validation results by using the data from commercial scale 2-KGA cultivation show good generalization performance and noise tolerance of the prediction approach. According to the estimation of future medium volume by statistical analysis and the prediction of total product, 2-KGA concentration is predicted.Due to the risk of contamination and other disturbances, the abnormality of fermentation processes often arises. The on-line fault detection method based on the prediction of product formation is presented for industrial 2-KGA cultivation. The results demonstrate that the proposed method could rapidly detect abnormalities of the processes.2. On-line computation and prediction of the profit functionThe profit function is an integrated index to describe the cost-effect of the 2-KGA fermentation processes, which is defined as the gross profit of a batch over its production time. The profit function is online calculated according to the mass balance and further predicted with the SVM-based rolling learning-prediction technique, which is potentially applicable for optimal scheduling. In the dissertation, it is also discussed how to establish and update the historical training database. Two updating database methods are proposed according to the profit category and the K-NN algorithm, respectively. The prediction results using the data from industrial scale 2-KGA cultivation indicate the advantage of the K-NN algorithm-based updating method and that may be independent on the inoculation sequences. 3. Optimal scheduling for 2-KGA cultivationAn optimal scheduling approach for 2-KGA fermentation process is proposed to improve allocation of L-sorbose resource with the aim of maximizing the economic profit of the multi-bioreactor workshop. The empirical operation in 2-KGA cultivation under study is to assign the same quantity of L-sorbose to each batch regardless of the batch-to-batch variations, while the optimal scheduling approach presented will determine the L-sorbose feeding according to the evaluation of the profit-making ability of the individual batch. Each batch is on-line evaluated based on the batch classification and its scheduling function. The batches of high profit-making ability will be fed more L-sorbose to lengthen the cultivation of these batches and exploit their profit contribution, while the poorly performed batches will be decreased the quantity of fed L-sorbose and terminated earlier to avoid profit loss. As a result, the scheduling strategy will make use of the L-sorbose resource more effectively and yield higher overall profit. Pseudo-on-line scheduling is carried out by using the data of industrial 2-KGA batches. The total profit increase of ca. 7% is demonstrated in comparison with the empirical operation.4. Developing on-line prediction and optimal scheduling software for 2-KGA fermentation processesFinally, the software is developed for 2-KGA fermentation processes based on the theoretic researches in this dissertation. The software real-time collects data from 2-KGA industrial cultivation. It is able to realize on-line prediction of the key state variables such as total product and product concentration, computation and prediction of the profit function, on-line evaluating the profit-making potential of the current batches, and optimal scheduling for multi-bioreactor system. The software has been installed and run successfully in the 2-KGA fermentation workshop.更多还原