【作者】 朱燕飞；

【导师】 毛宗源； 田联房；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 本文系统地分析了锌钡白回转窑煅烧过程的复杂特性,在过程采样数据分析的基础上,详细地探讨了该过程多种智能建模算法的理论及应用方法。文章首先结合国内外对回转窑煅烧过程建模及控制的研究现状,从回转窑煅烧生产过程特性出发,分析过程建模需待解决的难点及重点,提出过程分段建模的思想,为后续建模研究的展开奠定基础。其次,针对窑头温度控制系统的闭环辨识问题,应用两阶段闭环辨识方法对其展开研究。系统地分析了该方法克服闭环系统输出信号通过反馈环节与输入信号相关而对系统辨识造成的影响,通过仿真分析,建立了窑头温度随回油阀开度控制量的线性模型,并运用自相关函数方法检验模型的一致无偏性。针对回转窑煅烧段过程质量控制系统的建模问题,从基于能量平衡的控制思想出发,即在稳定窑头温度、物料流量和物料干燥效果的前提下,调节煅烧转速,以此来改变煅烧时间,调节过程反应的能量值,改善消色力指标。依据阿累尼乌斯经验方程推导过程煅烧段能量平衡控制的核心思想。在此基础上,建立了过程煅烧转速对数与煅烧温度倒数的线性回归预测模型,并对其模型的特性及逼近精度进行了分析和讨论。然后,为提高回转窑煅烧段控制模型建立的精度,修改了传统的基于煅烧机理的建模方法, 将模糊规则和神经网络结合起来, 提出了一种基于T-S (Takagi-Sugeno)模型的自适应神经模糊推理系统(ANFIS)建模方法。它采用T-S 的模糊辨识模型,运用神经网络为模糊模型的结构辨识和参数辨识提供自适应学习功能,较基于能量平衡的线性回归建模方法,在辨识精度上有很大的提高。在数据聚类算法研究的基础上,提出采用基于人工免疫系统(AIS)的数据聚类方法,解决ANFIS 网络的模糊结构辨识问题。它使网络能快速、灵活的调整其模糊规则的结构,在数据量大、工况复杂的过程辨识中有较强的实用价值。文章深入分析了AIS 网络中抑制阈值和聚类范围比例对系统辨识效果产生的影响,针对AIS 的随机性问题,对算法做了合理的修正,防止其造成聚类规则数的大幅波动。为提高回转窑煅烧段控制模型的辨识速度,文章提出了基于最小二乘支持向量机(LS-SVM)的建模算法。这种采用统计学习理论,基于结构风险最小化原则进行过程建模的思想,是解决复杂非线性系统辨识问题又一新的尝试。LS-SVM 采用最小二乘线性系统代替SVM 用二次规划方法实现学习问题,其结构简单,算法简练,在精度要求范围内,它有更优良的学习速度。通过仿真,得出其较ANFIS更好的辨识精度和速度。在提高过程模型特性的识别能力上,文章分析了两种典更多还原

【Abstract】 In this paper, the complexity of the calcination process of rotary Lithopone kiln is analyzed. On the basis of process data acquisition and analysis, several intelligent modeling methods for process control have been presented and discussed in detail. Firstly, after studying the present state of the modeling and control of calcination process of rotary kiln in domestic and foreign countries and analyzing the difficulties and key problems to be resolved urgently, segmentation modeling strategy is proposed, which establishes the foundation for subsequent modelings. Secondly, a two-stage identification method is proposed for the identification of the temperature control system of kiln head, which can overcome the shortcomings resulted from the correlation between the feedback and input. Moreover, the linear model of the temperature of kiln head as a function of the jaw opening of oil return valve is established through simulation For the modeling of the quality control system of the calcination process of rotary kiln, based on the idea of energy balance, under the condition of stabilizing the temperature of kiln head and flow rate and dry result, a new method is proposed to adjust calcination speed so that changing calcination time and adjusting the energy value of the calcination process and changing the ACC index. In addition, energy balance control of calcination process is deduced using Arrhenius empirical equation, on this basis, the linear regression prediction model concerning the logarithm of calcination angular speed versus the calcination temperature is established. Furthermore, the characteristics and approximation accuracy of the model are also discussed Thirdly, in order to improve the accuracy of the model of the calcination temperature of rotary kiln, a new modeling method—adaptive neuro-fuzzy inference modeling system (ANFIS) based on T-S model is proposed combining fuzzy logic with neural networks. By employing T-S identification model and using the learning ability of the neural networks, it can greatly improve the identification accuracy compared to the traditional linear regression modeling method. For the study of data clustering method, a novel clustering method based on artificial immune system (AIS) is developed to solve the problem of fuzzy structure identification, which makes the adjustment of fuzzy rules fast and flexible. This appears very useful in the process control with huge data and complex environment. In this paper, the influence on the system identification result by the suppress threshold and clustering range ratio in AIS network is also discussed in detail. Considering the randomness of AIS, the algorithm is modified to prevent the rule number of clustering from fluctuation In order to enhance the identification speed of the control model of the rotary calcination kiln, a novel least square support vector machines (LS-SVM) is proposed, which is another new try for solving the problem of complex system by employing statistical learning theory and establishing process model based on the principle of structure risk minimization. LS-SVM applys least squares linear system to replace the quadratic programming algorithm to realize its learning function, which has a simple structure, is easy of use and has an excellent learning speed within required accuracy. Through simulations it demonstrates more better identification accuracy and faster speed compared to ANFIS. In enhancing the identification capability of the proposed algorithm, a new modeling algorithm based on mixed kernel function is developed after analyzing the mapping of two typical kernal functions, which synthesizes the merits of suppressing prediction output fluctuation of global kernal function and the higher fitting accuracy of local kernal function and thus has excellent performance of synthesized identification compared to the SVM with single kernal function. Finally, the modeling strategy of LS-SVM is applied to design a multiple inputs and single output system, in which a model with multiple variables is establis更多还原