【作者】 徐飞；

【导师】 刘文琦；

【作者基本信息】 大连理工大学 ， 控制工程， 2012， 硕士

【摘要】 采用建立发酵模型来研究微生物发酵工艺是近年来研究工作的热点。发酵模型能够快速寻找到重组蛋白质药物适合的发酵工艺,并且能够对生产过程进行监控,促进重组蛋白质药物的研发和产业化生产。因此本文旨在：针对蛋白质发酵过程建立一个预测精度高并且泛化能力强的发酵模型以期优化发酵工艺、提高发酵生产质量。由于发酵过程的机理模型难以确立,目前常采用黑箱建模的方法针对发酵过程进行建模。最小二乘支持向量机(Least Squares Support Vector Machine, LS-SVM)方法改善了神经网络方法存在的对训练样本外的数据预测能力差、过学习、容易陷入局部极小点等方面的不足；同时相对于支持向量机(Support Vector Machine, SVM)而言,在损失了一定建模精度的前提下,原理更为简单、建模时间也大大缩短了,因而最小二乘支持向量机建模具有建模速度快、预测性能良好和学习能力强等优点。最小二乘支持向量机存在缺少稀疏性和鲁棒性的问题,本文首先采用剪枝法和加权法对最小二乘支持向量机进行了改进。由于对最小二乘支持向量机而言核函数的选择和构造是非常重要的,因此本文接下来针对核函数进行了研究,列出了几种单核函数,并从机理角度和实际应用效果两方面综合评价每个核函数。核函数分为全局性核函数(即旋转不变核)和局部性核函数(即平移不变核),全局核函数的学习能力较弱但泛化能力强,而局部性核函数的泛化能力较弱但学习能力强,本文针对核函数的这种特性提出了将全局核与局部核线性组合的混合核函数,克服了单核最小二乘支持向量机建模的缺点,并通过权重因子调节全局核与局部核对混合核函数的作用,这样能够使所建立的混合核函数最小二乘支持向量机模型具有较好的泛化能力和较高的预测精度。之后,为了提高所建立模型的精度,利用量子粒子群算法(QPSO)对核参数、惩罚系数和混合核函数中的权重因子进行优化调整。本文将基于混合核函数的最小二乘支持向量机建模方法应用于蛋白质发酵过程建模中,用量子粒子群算法对参数进行寻优,从仿真结果可以看出：混合核函数最小二乘支持向量机模型与单一核函数最小二乘支持向量机模型相比,学习能力和泛化能力均有所提高。用混合核函数方法建立的蛋白质发酵过程模型具有较好的模型拟合精度和满意的预测效果。更多还原

【Abstract】 Using fermentation model for the microorganism fermentation process and production process has been a hot topic for researchers in this area. The fermentation process model can find the suitable fermentation process for the protein drugs very quickly, and can monitor the production process. So it can promote the research and industrial production of protein drugs. The purpose of this paper is to establish an excellent and effective fermentation model with high accuracy of predictive capability for the protein fermentation process.It is very difficult to establish mechanism model. So we adopt the black-box modeling method. Least squares support vector machine (LS-SVM) method improves the neural network method that has easily into the local minimum points and poor ability to the training data out of the sample, but also improves the support vector machine (SVM) method that principle is too complex and takes a long time to modeling. The least square support vector machine modeling method has fast learning speed, good tracking performance and good regression accuracy on small sample model. So this paper uses the LS-SVM modeling method that suits small sample. We improve the sparsity and robustness of LS-SVM. In LS-SVM’s theory, choosing and constructing a suitable kernel is very important. So this paper firstly researches on the kernel and lists several single kernels. We evaluate each kernel from the mechanism point and the effect of practical application. The kernel is divided into global kernel and local kernel. The global kernel has generalization performance but weak learning ability. The local kernel has weak generalization performance but strong learning ability. This paper presents a multiple kernels function was linearly combined by global kernel and local kernel, and using it in the LS-SVM model. Multiple kernels method overcomes the shortcomings of the single kernel method. Then we optimize the mixed weight coefficient and the hyper parameter by using QPSO. The influence of the different kinds of kernel can be tuned by weight coefficient. Thus the model can get a better result on the regression accuracy and the generalization performance.In this paper, we will use the multiple kernels method to establish the LS-SVM fermentation model. This method applied in the microorganism fermentation modeling. We use QPSO to optimize the parameters. We can see that the LS-SVM model with multiple kernels has strong learning ability and generalization performance than the LS-SVM model with single kernel. The application results show that:the LS-SVM protein fermentation process model with multiple kernels has high accuracy and satisfactory prediction effect.更多还原