【作者】 公彦杰；

【导师】 高学金；

【作者基本信息】 北京工业大学 ， 控制科学与工程， 2010， 硕士

【摘要】 微生物发酵涉及到制药、食品等多个工业领域,与经济发展和人民生活密切相关。高成本和高能耗是微生物发酵生产的特征,为了提高发酵单位,降低生产成本,实现对微生物发酵过程的优化控制就成为了一个重要课题。由于微生物发酵过程是一类非常复杂的生化反应过程,人类尚未完全弄清楚它的机理。并且现有的在线生物传感器的测量精度难以保证,生物参数主要通过离线分析得到,这往往存在较大的滞后,无法及时反馈控制信息。所以,建立高精度的发酵过程产物预估模型,就成为优化控制需要研究的核心内容。本文针对发酵过程时变性、非线性、不可逆、多变量耦合等特点,通过对现有发酵过程建模方法的对比研究,提出了基于动态时间弯曲距离(DTW)的最小二乘支持向量机(LS_SVM)在线建立发酵过程局部模型的方法。主要有以下几个方面的研究成果:1、基于DTW的在线构造相似训练样本集的方法:该方法首先将当前批次滑动时间窗内数据作为查询序列,以DTW作为判断时间序列相似性的标准,从历史批次数据库中搜索与之相似度最高的数据区间,组成在线训练样本集。2、模型输入变量的选取方法以及超参数敏感度分析:通过仿真实验,分析了不同输入变量以及核函数对模型均方误差(MSE)的影响,选择了适合发酵过程的RBF核函数,分析了模型精度对γ、σ2的敏感程度,确定了最优超参数的取值范围。3、基于粒子群交叉验证(PSO-CV)的在线超参数优化方法:通过对交叉验证确定模型超参数方法的分析,提出了以最小化K-CV均方误差为目标的PSO超参数优化算法,在保证模型精度的情况下,兼顾了模型的泛化能力,与网格搜索法相比,有更好的性能。4、在线建模软件的开发:利用VC++6.0开发了Windows系统下的发酵过程在线建模软件,实现了通过OPC方式读取组态软件中新采集到的数据,通过ADO方式读取数据库中历史批次的数据,对实际发酵过程建立基于DTW的LS_SVM在线局部模型,同时绘制模型预估输出以及主要可测变量的动态实时曲线,有助于实现发酵过程的优化控制。更多还原

【Abstract】 The microbial fermentation involves the pharmaceutical, food and other industries, and is closely related to economic development and people’s everyday life. High cost and energy consumption are the characteristics of microbial fermentation. Therefore, in order to improve the fermentation unit and reduce production cost, it has become an important issue to realize optimal control of fermentation process.Since the fermentation process is a kind of very complex biochemical reaction process, the human have not fully made clear of its mechanism yet. What is more, the existing online biosensor is difficult to ensure measurement accuracy, the measurement of biomechanical parameters is mainly obtained through off-line analysis, it usually has a big lag which can’t feedback control information in time. Therefore, the establishment of high-precision prediction model of fermentation process products becomes core content to be researched in optimal control.In this paper, we aim at the characteristics of fermentation process, including variability、nonlinear、irreversible and multivariable coupling. By comparing existing fermentation process modeling methods, a new online local modeling method is proposed for fed-batch fermentation processes based on dynamic time warping (DTW) and least squares support vector machine (LS_SVM). The major research findings and innovation are as follows:1) Online constructing similar training sample set based on DTW:A set of data within the sliding window is set as a query sequence in the current batch, and then using DTW as standard to judge the similarity of time series, we search for the most similar sub-sequence from the historical batch database to form the training set.2) Selection method of model input variables and parameters sensitivity analysis:By analyzing the influence on model MSE of different input variables and the kernel function through simulation experiments, I select RBF kernel function for the fermentation process and analyze sensitivity degree of the model accuracy onγandσ2, and determine the optimal range of super parameters.3) Online super parameters optimization method based on PSO-CV:Through analysis of cross-validation method used to determine the model’s super parameters, a new online super parameters optimization method based on Particle Swarm Optimization(PSO)is proposed to minimize the K-CV error of for the model, it not only ensure accuracy of the model, but also take into account the model generalization ability. Results show that PSO-CV method has better performance compared with the grid search method.4) Online modeling software development:A set of software for modeling and optimization of the fermentation process has been designed by VC++ 6.0 in Windows system. The software can read current and historical data through OPC and ADO to establish the DTW LS_SVM online local model of actual fermentation process, at the same time the dynamic real-time curve of model predictive output and the major measurable variables can be draw when the new data arriving. The development of the software is helpful to realize the optimal control of the fermentation process.更多还原