【作者】 王雅琳；

【导师】 桂卫华；

【作者基本信息】 中南大学 ， 控制理论与控制工程， 2001， 博士

【摘要】 在工业过程自动化方面，稳态优化控制已成为连续生产企业挖掘潜力、降低成本、增产增效、提高竞争能力的重要途径。但是，由于工业过程反应机理复杂，非线性、时变、强耦合严重，建立生产目标与操作参数间的稳态关系模型非常困难，从而影响了复杂工业过程优化控制的实现。因此，为实现复杂工业过程的稳态优化控制，本文重点研究了工业过程的智能集成建模方法及其工业应用。 首先，论文系统地分析和总结了过程控制、稳态优化控制、工业过程模型化研究以及有色金属冶炼过程控制的研究现状以及存在的问题，针对现有建模方法存在的缺陷、复杂工业过程的特点以及优化控制对模型的要求，提出智能集成建模的思想，较全面和系统地建立了智能集成建模理论的基本框架，即给出了智能集成建模的一般定义和六种集成形式：并联补式集成、加权并集成、串联集成、模型嵌套式集成、结构网络化集成以及部分方法替代式集成；同时，还给出了智能集成建模的形式化描述以及工程实现。该理论框架为新的建模方法的派生提供指导。 接着，在所提出的智能集成建模理论框架的指导下，根据NN的研究现状以及在工业过程建模中存在的问题，以神经网络建模方法为基础探讨了三类智能集成方法，即（a）多个神经网络之间的集成；（b）神经网络与传统建模方法间的集成；（c）神经网络与其它智能技术（如模糊技术）间的集成。第一类集成中提出了一种自适应的监督式分布神经网络ASDNN建模方法。该方法按照监督式聚类方法将学习样本集划分为多个具有不同特征值的子空间，然后建立一组神经网络模型，并按模糊组合方式将多个NN连接。ASDNN模型克服了单个神经网络在工业过程建模中出现的结构庞大、收敛慢、过拟合等问题，并且监督式聚类不需预先设定样本集的聚类中心个数，同时，保证了基于数据的样本空间划分的有效性和合理性；另外，ASDNN采用不同自适应学习机制实时修正网络模型，进一步提高了模型的实时预测能力。第二类给出了嵌套式、并集补以及结构网络化三种传统方法与神经网络集成建模的形式，并提出了一种用最小二乘法对神经网络输入加权的集成建模方法，该方法预测精度高，收敛速度较快，适合于非线性严重，特别是输入变量较多的工业过程。第三类集成中主要研究了一种模糊系统的网络化集成建模方法。该方法以ANFIS网络五层结构实现Sugeno 0型和Sugeno Ⅰ型模糊模型的网络化；基于已有的专家知识或采用竞争学习聚类方法初始化网络前件参数同时确定最初网络结构；在初始网络结构保证了网络学习可免于陷入局部极值的前提下，采用提出的加动量项模糊增强型BP算法和最小二乘混合学习算法提高学习速度，并用赤池信息准则AIC确定最优的网络结构；最后针对工业过程的时变性，还增加了网络参数的递推修正以及结构的增长式修正。用该方法建模，网络结构简单，易于学习，能较好实现模糊推理，可避免陷于局部极值点。 工业应用部分，论文以有色冶炼过程的优化控制为背景，以工业过程模型化研究为重点，探讨了锌电解过程分时供电优化调度和铅锌火法冶炼过程的建模与优化控制问题。针 王雅琳：智能集成建模理论及其在有色冶炼过程优化控制中的应用研究对具体的研究对象，进行了多种智能集成建模方法及其工业应用的研究，包括：昨）锌电解过程电流效率与电流密度和酸锌比间的模糊神经网络模型；O）基于经验机理模型和ASDNN的烧结块铅锌成分智能集成模型；k）由基于物料平衡和神经网络的数学模型和基于专家经验的规则模型并联组成的烧结块殁硫软测量集成模型；N）基于最优权重的烧穿点位置与透气性和台车速度的模糊规则与神经网络并联集成模型；忙）鼓风透气性与批重和批次的输入加权式神经网络模型。工业应用结果证明了这些智能集成建模方法的可行性、优越性和实用性。文中还在建立了过程优化模型的基础上，分别采用模拟退火算法、专家推理和遗传算法求解了三个不同的过程优化问题。其中，锌电解过程分时供电优化调度系统己投入生产运行，为锌电解过程带来显著的经济效益；而铅锌火法冶炼过程的优化控制只给出部分优化仿真结果和优化控制思路，尽管如此，他们验证了智能集成模型用于过程优化中的可行性，并能为控制的具体实施提供重耍指导。更多还原

【Abstract】 Nowadays, steady-state optimal control has become an important approach to tapping potential, reducing cost, increasing production, raising efficiency and improving competitive power of processing industrial enterprise. However, owing to complex process mechanism with strong non-linearity, time-variability and severe coupling, it is difficult to construct complex steady-state relationships between production objects and operational parameters, which obstructs the implement of optimal control for complex industrial process. So this dissertation emphasizes on intelligent integrated modeling methods and their industrial applications to solve the above problems.Firstly, the dissertation systematically analyzes and summarizes the status quo and existing problems of researches on process control, steady-state optimal control, industrial process modeling and non-ferrous metallurgical process control. Then aiming at the disadvantages of the present modeling methods, the characteristics of complex industrial process and the requests of optimal control for models, an idea of intelligent integrated modeling is proposed and the basic framework of the corresponding theory is constructed comprehensively and systematically. Namely, its universal definition and six integrated forms including parallel connection for compensation, parallel connection with weight, model integration in series, model embedding, neural networks implement of model structure and part substitution of modeling are described, and its formalized description and engineering implement are given. The constructed modeling theory can provide guidance for derivation of new modeling.Secondly, according to research status on NN(Neural Networks) and existing problem in modeling for industrial process, three integrated modeling based on NN are studied under the instruction of the intelligent integrated modeling theory. In Multi-NN as the first integrated modeling, an adaptive supervised distributed NN (ASDNN) is proposed. According to ASDNN, learning sample set is divided into many subsets with different eigenvalue by supervised clustering method, a series of NN models are constructed based on these subsets, then these NN models are combined by fuzzy logic technology. ASDNN has many advantages: 1) solving some problems in NN for process modeling such as large structure, slow convergence as well as overfitting and so on; 2) directly determining clustering numbers without pre-determination, guaranteeing efficient and reasonable division of sampling set; 3) strengthening its real-time predictability. The second integration modeling involves the integration of NN and the conventional modeling. At first, model embedding, parallel connection for compensation and neural networks implement of model structure the three integration forms are described in brief. Then NN integration modeling with weighted input layer by least-square Methods(LSM) isproposed, which has not only high accuracy and fast convergence but also excellent adaptability to the industrial process with strong non-linearity, especially many input variables. The third integration, integrating NN and other intelligent modeling, emphasizes on a fuzzy system integrated modeling with NN structure. The modeling has some characteristics as follow. I) describing fuzzy models of Sugeno 0 and Sugeno I as the five-layer structure of ANFIS; 2) initiating both antecedent parameters and structure based on the present expertise or competitive learning clustering algorithm to avoid local minimum; 3) proposing a hybrid algorithm combined LSM and BP algorithm with changeable learning step and momentum factor altered by fuzzy logic to speed up network learning, and optimizing network structure based on Akaike Information Criterion(AIC); 4) real time updating network parameters and structure for timevariability of industrial process. The modeling due to its simple structure and learning facility can execute fuzzy reasoning better and avoid local minimum.As far as industrial applications of the intelligent integrat更多还原