【作者】 蒋华琴；

【导师】 刘兴高；

【作者基本信息】 浙江大学 ， 工业过程统计建模和优化， 2012， 硕士

【摘要】 熔融指数(MI)是聚丙烯生产的重要指标之一,建立可靠的熔融指数预报模型非常重要。丙烯聚合过程机理复杂,生产工艺多样,设备工段繁多,因此采用机理建模方法存在相当大的难度。而统计建模方法是一种依赖数据的方法,对系统内部机理的了解要求很少,所以在丙烯聚合熔融指数预报中得到广泛应用。统计学习理论是一种基于小样本的机器学习理论,支持向量机是在此理论基础上提出来的。它根据结构风险最小化原则,通过核函数在一个高维特征空间中构造线性决策函数,避免了维数灾难,且可达到全局最优解。支持向量机良好的性能使其成为机器学习领域的热点课题。支持向量机的性能依赖其参数的选择,本文应用智能优化算法进行参数寻优,从而建立多种智能支持向量机模型。本文的主要内容包括：1.为了降低标准支持向量机(SVM)的计算复杂度,提高其学习速度、泛化能力和稀疏性,本文研究了最小二乘支持向量机(LSSVM).加权最小二乘支持向量机(WLSSVM)和相关向量机(RVM),并用于熔融指数预报。实验结果表明上述方法的可行性和有效性。2.核函数参数和惩罚参数决定了支持向量机的性能,选择最佳的参数可以直接提高模型预报能力。针对支持向量机的参数选择问题,本文采用带有权重因子的改进粒子群优化算法,分别对LSSVM、WLSSVM和RVM的参数进行寻优,建立了单纯智能支持向量机模型(PSO-LSSVM、PSO-WLSSVM和PSO-RVM)。PSO算法具有很强的寻优能力和快速的收敛速度,能在最短的时间内找到函数的全局最优点,使参数寻优成为可能。实验结果表明优化后的模型具有更好的预报效果。3.针对标准粒子群算法在迭代过程中易出现粒子过早收敛而陷入局部最优的缺陷,通过引入免疫系统的抗体选择机制,构造了一种基于免疫机制的免疫粒子群优化(IC-PSO)算法,来保持更新粒子的多样性,从而克服标准粒子群算法过早收敛的缺陷；为了减小粒子群搜索的盲目性,避免早熟,本文利用蚁群算法为免疫粒子群算法找到一条最优路径,构造了蚁群-免疫粒子群优化(AC-ICPSO)算法。然后利用这两种优化方法对LSSVM和WLSSVM进行参数寻优,建立了混合智能支持向量机模型(ICPSO-LSSVM、AC-ICPSO-LSSVM. ICPSO-WLSSVM、AC-ICPSO-WLSSVM).以实际聚丙烯生产的熔融指数预报作为实例进行研究,结果表明所提出模型的有效性和良好的预报精度。4.针对丙烯聚合生产控制中,系统存在高复杂性、不可确定性、多层次性等特点,本文提出了模型在线校正策略。随着数据的更新,不断调整预报模型以适应最新工况。实验表明,校正后的模型预报效果更好。更多还原

【Abstract】 Melt index (MI) is considered as one of the important variables of the quality in the propylene (PP) polymerization process, and it is crucial to propose a reliable estimation model of MI.. PP polymerization process usually involves complex kinetic mechanism, and various plants, which makes it a challenge to study the process though mechanism modeling approaches. Based on the dada, the statistical method doesn’t need much knowledge of the studied process, so it is used to predict MI popularly.Statistical Learning Theory (SLT) is a theatrical framework of machine learning for small samples, and Support Vector Machine (SVM) comes out from this theory. According to the structural risk minimization rule, it can get the global optimal linear decision function in a higher dimensional feature space according to a kernel function. It avoids the curse of dimensionality and it is of good generalization ability. Its good performance makes it become a hot topic in machine learning. However, its good performance depends on the chosen of parameters, and this paper makes the use of intelligent optimization algorithms to optimize the parameters, so many kinds of intelligent SVM models are proposed. The main contributions of the present work are as follows:1. In order to improve the training speed、generalization ability and space ability of the traditional SVM, this paper does some research on Least Square SVM (LSSVM)、Weighted LSSVM (WLSSVM)、Relevance Vector Machine (RVM), and these methods are used to predict MI. The results confirm the models’validity.2. The kernel parameter and the regularization factor determine the performance of SVM, so obtaining the best parameters can improve the prediction ability of models directly. In order to choose better parameter for SVM,this paper applies improved Particle Swarm Optimization (PSO) algorithm, which has an inertia weight factor, to optimize the parameters of LSSVM. WLSSVM and RVM, so the pure intelligent SVM models (PSO-LSSVM. PSO-WLSSVM and PSO-RVM) are proposed. The good search capability and fast convergence make PSO algorithm applicable to parameter optimization. The results show that the optimal models have better prediction ability.3. Addressing the deficiency of the Particle Swarm Optimization (PSO) algorithm whose particles are easy to sink into premature convergence and run into local optimization in the iterative process, with the clone selection strategy of immune system, the Immune Clone PSO (ICPSO) algorithm based on immune strategy was proposed to make the particles of ICPSO maintain the diversity during the iterative process so as to overcome the defect of premature convergence of PSO; in order to reduce the blindness of the research of PSO and avoid premature, Ant Colony Optimization (ACO) algorithm is used to find an optimal path for ICPSO algorithm, thus the AC-ICPSO algorithm is proposed. Then the two algorithms are used to optimize the LSSVM and WLSSVM’s parameters, so the hybrid intelligent SVM models (ICPSO-LSSVM、AC-ICPSO-LSSVM、ICPSO-WLSSVM. AC-ICPSO-WLSSVM) are found. Researches on the optimized model were illustrated with the real plant of propylene polymerization, and the results showed that the proposed approach had great prediction accuracy and validity.4. In order to avoid the characteristics of high complexity, uncertainty, multi level nature and so on in PP polymerization process, Online Correction Strategy (OCS) is presented in this paper. With the update of the data, the forecast models need adjustment constantly to adapt the latest conditions. The experiments show that the corrected models have better prediction ability.更多还原